Artificial Intelligence at the Jet Propulsion Laboratory

Home , New Millennium program, Palomar Observatory

AI Magazine Volume 18 Number 1 (1997) (© AAAI) Articles Making an Impact Artiﬁcial Intelligence at the Jet Propulsion Laboratory

Steve Chien, Dennis DeCoste, Richard Doyle, and Paul Stolorz

■ The National Aeronautics and Space Administra- described here is in the context of the remote- tion (NASA) is being challenged to perform more agent autonomy technology experiment that frequent and intensive space-exploration mis- will fly on the New Millennium Deep Space sions at greatly reduced cost. Nowhere is this One Mission in 1998 (a collaborative effort challenge more acute than among robotic plane- involving JPL and NASA Ames). Many of the tary exploration missions that the Jet Propulsion AI technologists who work at NASA expected Laboratory (JPL) conducts for NASA. This article describes recent and ongoing work on spacecraft to have the opportunity to build an intelli- autonomy and ground systems that builds on a gent spacecraft at some point in their careers; legacy of existing success at JPL applying AI tech- we are surprised and delighted that it has niques to challenging computational problems in come this early. planning and scheduling, real-time monitoring By the year 2000, we expect to demonstrate and control, scientific data analysis, and design NASA spacecraft possessing on-board automat- automation. ed goal-level closed-loop control in the planning and scheduling of activities to achieve mission goals, maneuvering and pointing to execute these activities, and detecting and I research and technology development resolving of faults to continue the mission reached critical mass at the Jet Propul- without requiring ground support. At this Asion Laboratory (JPL) about five years point, mission accomplishment can begin to ago. In the last three years, the effort has become largely autonomous, and dramatic begun to bear fruit in the form of numerous cost savings can be achieved in the form of JPL and National Aeronautics and Space reduced, shared ground staffing that responds Administration (NASA) applications of AI on demand to beacon-based requests for inter- technology in the areas of planning and action originating from the spacecraft. Indeed, scheduling, real-time monitoring and control, a New Millennium Program Mission Opera- scientific data analysis, and design automa- tions Study estimated that remote-agent tech- tion. Such successes, described in detail in nology could reduce mission operations cost, this article, have also set the stage for JPL AI exclusive of data analysis, by as much as 60 researchers and technologists to seize an percent. unprecedented opportunity: the commitment By 2005, we expect that a significant por- by NASA to the development of software tion of the information routinely returned technology to realize highly autonomous from space platforms would not be raw data, space platforms. This strategic shift by NASA, and would not simply and strictly match fea- in which AI technology will play a central tures of stated prior interest, but would be and critical role, includes an important his- deemed by the on-board software to be inter- torically missing piece of the picture: the esting and worthy of further examination by availability of space missions whose primary scientists on the ground. At this point, limit- purpose is the validation of new technologies. ed-communications bandwidth would be used NASA’s New Millennium Program fills exactly in an extremely efficient fashion, and science this gap. Some of the most important work alerts from various and far-flung platforms

would be anticipated with great interest. data from digitized photographic plates of The first steps toward realizing this vision the night sky collected at the Mt. Palomar are happening now and are described in the astronomical observatory-identifying and following paragraphs. Although the early goal classifying sky objects indexed into a compre- for autonomy technology is the reduction of hensive catalog containing approximately mission operations costs, the ultimate payoff three billion entries. SKICAT was able to classify will be the enabling of new mission classes extremely faint objects that were beyond the and the launching of a new era of solar sys- reach of expert-astronomer visual inspection. tem exploration, beyond reconnaissance. It was recently utilized to discover 16 new Spacecraft missions in this new era will be quasars, among the most distant objects in characterized by sustained presence and in- the universe. For the first time, astronomers depth scientific studies performed by free have an objective basis for conducting flyers, orbiters, and ground vehicles, some- unprecedented, large-scale cosmological stud- … mission times arrayed in constellations. Autonomy ies. SKICAT remains one of the most outstand- will be the central capability for enabling ing successes for machine-learning technolo- accomplish- long-term scientific studies of a decade or gy to date. This work was led by Usama ment can more, currently prohibited by cost, and Fayyad, in collaboration with George Djor- enabling new classes of missions that inher- govski of the Astronomy Department at the begin to ently must be executed without the benefit of California Institute of Technology (Caltech). become ground support, either because of control The MVP system (Chien and Mortensen challenges, for example, small-body (asteroid 1996) applied AI planning techniques to the largely and comet) rendezvous and landing missions problem of automatic software configuration autonomous, or because of the impossibility of communi- for image analysis. The VICAR set of image- and dramatic cation for extended periods, for example, an processing routines has been developed over underice explorer at Europa or a Titan aer- a period of many years at JPL. These routines cost savings obot. The vision for future NASA missions support image-processing steps such as can be based on intelligent space platforms is mosaicking and color-triplet reconstruction. tremendously exciting. Powerful but cumbersome to use, the VICAR achieved in The need for autonomy technology is routines support an essential processing the form of nowhere greater than in the set of deep space phase before true scientific image analysis can planetary missions that JPL conducts for begin. MVP is a front end for VICAR and reduced, NASA. The extreme remoteness of the targets, allows scientists and other users to simply shared ground the impossibility of hands-on troubleshoot- state their image-analysis goals. The system staffing that ing or maintenance, and the difficulties of automatically analyzes data dependencies light-time delayed communication (over four and other constraints and configures the responds on hours round trip to the outer solar system) all appropriate VICAR routines to achieve these demand to contribute to make JPL science missions the goals. MVP reduces the time to construct a focus of the development and application of typical VICAR job from 4 hours to 15 min- beacon-based autonomy technology (Doyle, Gor, et al. utes for expert users and from days to hours requests for 1997). JPL has been designated the lead NASA for novice users. The system is being used to center for spacecraft autonomy not only support the analysis of images being returned interaction because of the nature of its missions but also from the Galileo spacecraft during its current originating because of its unique combination of resident tour of the Jupiter planetary system. This from the expertise in AI, spacecraft engineering, space work was led by Steve Chien, in collaboration mission design, and systems engineering. with Helen Mortensen of JPL’s Multimission spacecraft. Not surprisingly, the imperative of cost Image-Processing Laboratory. constraints as drivers for the development of AI research and development activities at autonomy capabilities are balanced against JPL are conducted primarily by three research significant perceived risk in on-board uses of groups: (1) the Artificial Intelligence (AI) AI technology. An important ingredient in Group, led by Steve Chien, which focuses on making this opportunity credible and real are automated planning and scheduling and the previous successes at JPL in the applica- design automation; (2) the Machine Learning tions of AI. Two of the most notable of these Systems (MLS) Group, led by Paul Stolorz; successful applications are the sky-image cata- and (3) the Monitoring and Diagnosis Tech- loging and analysis tool (SKICAT) and the mul- nology (MDT) Group, led by Dennis DeCoste. timission VICAR (video image communica- The effort was once housed in the single AI tion and retrieval) planner (MVP). SKICAT group but has steadily grown: Its 30 to 40 (Fayyad, Djorgovski, and Weir 1996b) com- practitioners now make up 3 of the 9 groups pletely automated the process of reducing in JPL’s Information and Computing Tech-

104 AI MAGAZINE Articles nologies Research Section, led by Richard Planning, Scheduling, and Task Execu- Doyle. tion for Spacecraft Commanding In the remainder of the article, we describe Spacecraft command generation and valida- in greater detail the major AI projects at JPL. tion is an expensive and labor- and knowl- In particular, we describe efforts in the areas edge-intensive process. Enabling direct high- of planning and scheduling, monitoring and level commanding of spacecraft by diagnosis, knowledge discovery and data engineering and science personnel greatly mining, and automated design. reduces the requirements for highly skilled spacecraft-cognizant personnel during nomi- Planning and Scheduling nal operations, thereby cutting down on mission operations costs. A New Millennium Pro- As the NASA lead for unmanned exploration gram Mission Operations Study concluded of deep space, JPL has led missions to the out- that automating command and control func- er reaches of the solar system (as exemplified tions could have resulted in a savings of $14 by the Voyager, Galileo, and Cassini mis- million a year for a Magellan-class mapping sions). Although these missions have mission and $30 million a year for a Galileo- achieved enviable success, NASA is now being class multiple-flyby mission. challenged to perform future missions with In a ground-based context, direct com- smaller budgets, shorter cycle times, and manding of the spacecraft by science person- smaller science and operations teams. One nel also allows for the opportunity to conduct NASA is major element in mission operations is the truly interactive science—an embodiment of now being problem of command, control, and schedul- the concept of a virtual spacecraft on the ing. From an applications perspective, this internet. In certain cases, automated space- challenged area encompasses the determination of cor- craft commanding can enhance science return to perform rect actions and required resources for mis- by increasing the efficiency of resource man- sion operations, both for the spacecraft as agement (for example, data and power man- future well as for all the elements of the ground sys- agement). If it is possible to run the planner missions tem necessary to run the mission. From a on board the spacecraft, additional benefits with technology perspective, this area focuses on can be realized: First, communication with planning, scheduling, and task-execution the spacecraft can be reduced greatly in that smaller architectures. commands do not need to be uplinked, and budgets, In this section, we outline ongoing efforts reduced spacecraft state information can be at JPL in research, development, and deploy- downlinked. Second, by avoiding the commu- shorter ment of these technologies to automate com- nications loop, autonomous commanding of cycle times, mand, control, and resource-allocation func- the spacecraft with an on-board system allows tions to reduce operations teams, reduce for an immediate response to changes in and smaller command cycle times, and increase the spacecraft state (for example, faults) or discov- science and efficiency of the utilization of scarce resources, eries from science analysis. all targeted at enabling increased science and In a collaborative effort involving the AI operations space exploration at reduced cost. We begin group and the Sequencing Automation teams. by first describing efforts in the area of space- Research Group at JPL and the Computation- craft commanding and on-board task execu- al Sciences Division (Code IC) of the NASA tion, describing projects for the New Millenni- Ames Research Center, AI planning, schedul- um Deep Space One Mission and the Earth ing, and intelligent task control techniques Orbiter One Mission, the U.S. Navy UFO-1 are being developed and applied for on-board satellite, and the data-chaser shuttle payload. control of highly autonomous spacecraft. The We then describe projects in automation of automated scheduler and task-execution ground systems—specifically in automating technologies are being developed for the New operations of the deep space network (DSN), Millennium Deep Space One Mission, the which is used for communicating with space- first of the deep space missions planned for craft, navigating spacecraft, and using radio the New Millennium Program. This small science. We then describe the MVP Project to spacecraft will fly by an asteroid and a comet, use planning technology to assist in science with a launch expected in 1998. The primary data preparation and analysis. Finally, we objective of the New Millennium Program is describe basic technology work in machine the demonstration of new technologies that learning for next-generation planning and will greatly advance the state of the art in scheduling systems able to automatically space exploration. One of the new technolo- adapt to changing problem distributions and gies for the Deep Space One Mission is an context. autonomy software package, named the

SPRING 1997 105 Articles

remote agent, that will fly on board the In addition to the deployment of on-board spacecraft. The remote agent has three compo- planning and task control technology on the nents: (1) the smart executive, (2) mode iden- New Millennium Deep Space One Mission, tification and recovery, and (3) the planner- there are additional projects under way to scheduler; these components work together deploy planning and scheduling technology to autonomously control the spacecraft (Pell, in a ground-based context using the automat- Bernard, et al. 1996). ed scheduling and planning environment The planner-scheduler (Muscettola et al. (ASPEN) and also the data-chaser automated 1997) generates a sequence of low-level com- planning and scheduling system (DCAPS) for mands, given an initial state and a set of direct-science commanding. high-level goals from the scientists and engi- A number of successful applications of neers. Performing this task requires a automated planning and scheduling of space- significant knowledge of the spacecraft opera- craft operations have recently been reported tions, possible spacecraft states, operability in the literature. However, these applications constraints, and low-level commands that are have been one-of-a-kind applications that executable by the smart executive. In addi- required a substantial amount of develop- In addition to tion, heuristic knowledge about priorities and ment effort. To reduce this development deliberative preferences might be required to generate effort, the AI group at JPL has been working planning, an better-quality solutions in a shorter time. The on ASPEN (Fukunaga et al. 1997a), a modular, planner system builds a schedule while it reconfigurable application framework that is autonomous respects the encoded spacecraft constraints, capable of supporting a wide variety of plan- spacecraft science and engineering preferences, and syn- ning and scheduling applications. ASPEN pro- chronization with external processes. An vides a set of reusable software components requires the incremental refinement approach is used to that implement the elements commonly ability to provide an exhaustive search that guarantees found in complex planning-scheduling sys- the generation of a solution schedule from tems, including an expressive modeling lan- execute correctly encoded goals and knowledge. guage, a resource-management system, a tem- incompletely In addition to deliberative planning, an poral reasoning system, several search specified autonomous spacecraft requires the ability to engines, and a graphic user interface (GUI). execute incompletely specified plans and the The primary application area for ASPEN is plans and ability to respond quickly and intelligently to the spacecraft operations domain. Planning the ability unforeseen run-time contingencies. The MLS and scheduling spacecraft operations involves group, in collaboration with the Computa- generating a sequence of low-level spacecraft to respond tional Sciences Division of the NASA Ames commands from a set of high-level science quickly Research Center, is developing a smart execu- and engineering goals. ASPEN encodes com- tive to provide these capabilities for the New plex spacecraft operability constraints, flight and Millennium Deep Space One Mission (Pell, rules, spacecraft hardware models, science intelligently Gat, et al. 1996). experiment goals, and operations procedures to The executive is implemented using a lan- to allow for the automated generation of low- guage developed at JPL called the execution level spacecraft sequences by using con- unforeseen support language (ESL) (Gat 1997). ESL pro- straint-driven planning and scheduling tech- run-time vides a set of advanced control constructs that nology. ASPEN is currently being used in the simplify the job of writing code to manage development of automated planner-scheduler contingencies. multiple concurrent tasks in the face of unex- systems for commanding the UFO-1 naval pected contingencies. It is similar in spirit to communications satellite and the New Mil- RAPs (reactive action packages), RPL (reactive lennium Earth Orbiter One spacecraft as well plan language), and RS, and its design owes as a scheduler for ground maintenance of the much to these systems. Unlike its predecessors, highly reusable space transportation. ESL aims for a more utilitarian point in the Figure 1 shows information on the applica- design space. It was designed primarily to be a tion of ASPEN to generating operations plans powerful and easy-to-use tool, not to serve as a for the New Millennium Earth Orbiter One representation for automated reasoning or for- Mission. The top portion of figure 1 shows mal analysis (although nothing precludes its the interface to the scheduler, displaying the use for these purposes). ESL consists of several relevant observation activities (the observa- sets of loosely coupled features, including con- tion activities shown at the top on the ActTL tingency handling, concurrent task manage- timeline), the resources used (the observation ment, and a backtracking logical database. A instrument/ETM, data storage device/SSR, set of constructs for run-time resource manage- transponder, for example), and relevant ment is currently being developed. exogenous events (such as sun angle state).

106 AI MAGAZINE Articles

The bottom section of fiigure 1 shows a small portion of the temporal constraint network relating to the synchronization of a downlink activity with ground stations used to derive the schedule. The squares correspond to important events (activities, states, and resource values), and the brackets indicate the minimum and maximum time between the events. DCAPS (Rabideau et al. 1997) is a collaborative effort involving the AI group and the Sequencing Automation Research Group at JPL and the Colorado Space Grant Consor- tium at the University of Colorado at Boulder. In DCAPS, AI planning and scheduling techniques are being developed and applied to enable direct goal-based science and engineering commanding. Data-chaser is a shuttle payload scheduled for flight in July 1997. Data-chaser contains three solar science instruments and is part of the Hitchhiker Student Outreach Program. DCAPS uses iterative repair planning and scheduling techniques to automatically generate the low-level command sequence involving spacecraft operability constraints, science and engineering preferences, and synchronization constraints with external processes. The iterative repair approach to planning and scheduling is useful in that it allows for natural interaction with the user. Automated Planning and Scheduling for Operations of the Deep Space Network Each day at sites around the world, NASA’s DSN antennas and subsystems are used to Figure 1. ASPEN-Generated Plans for the New Millennium perform scores of tracks to support earth- Earth Orbiter One Spacecraft. orbiting and deep space missions (Chien, Top: Earth Orbiter One operations plan derived by the ASPEN scheduler. Bottom: Hill, et al. 1996; Hill, Chien, et al. 1996). Temporal constraint subnetwork used to derive temporal constraints on activi- However, the actual tracking activity is mere- ties relating to data transmission to a ground station. ly the culmination of a complex, knowledge- intensive process that actually begins years before a spacecraft’s launch. When the decision is made to fly a mission, a forecast is planning for DSN resources (Fox and Borden made of the DSN resources that the spacecraft 1994; Loyola 1993). will require. In the resource-allocation pro- As the time of the actual tracks approaches, cess, the types of service, frequency, and this estimate of resource loading is converted duration of the required tracks are deter- to an actual schedule, which becomes more mined as well as high-level (for example, antenna) resource requirements. Although concrete as time progresses. In this process, the exact timing of the tracks is not known, a specific project service requests and priorities set of automated forecasting tools is used to are matched up with available resources to estimate network load and assist in ensuring meet communications needs for earth-orbit- that adequate network resources will be avail- ing and deep space spacecraft. This schedul- able. The Operations Research Group has ing process involves considerations of thou- developed a family of systems that use opera- sands of possible tracks, tens of projects, tens tions research and probabilistic reasoning of antenna resources, and considerations of techniques to allow forecasting and capacity hundreds of subsystem configurations. In

SPRING 1997 107 Articles

because antennas are the central focus of ANT CMD TRK ETX UWV RCV TLM resource contention. After establishing a Start range of antenna options, DANS considers allo- Configure Configure Configure Configure Configure Configure Configure cation of the 5 to 13 subsystems (out of the antenna command MDA excitor microwave receiver telemetry tens of shared subsystems at each antenna controller processor controller controller group complex) that are used by each track. DANS Configure transmitter uses constraint-driven, branch-and-bound, best-first search to efficiently consider the Test translator on large set of possible subsystem schedules. Lock receiver Once a specific equipment and timing Configure SRA assignment has been made, there is the prob- for ranging lem of determining how to operate the equip- Calibrate CMD ment to achieve the requested services. channel delay Because of the complexity of the equipment, the large set of communications services (in CMD data validate and transfer the tens), and the large number of supported equipment configurations (in the hundreds), Configure exciter Configure receiver correctly and efficiently operating this equip- controller for track for track ment to fulfill tracking goals is a daunting task. CMD data validate and transfer Begin The DSN antenna operations planner acquisition (DPLAN) (Chien, Govindjee, et al. 1997, 1996) is an automated planning system developed by the AI group to automatically generate Figure 2. Plan Constructed by Deep Space Network Antenna Operations antenna-tracking plans to satisfy DSN service Planner for Precalibration of a 34-Meter Beam Waveguide Antenna for a requests. To generate these antenna opera- Telemetry Commanding and Ranging Track. tions plans, DPLAN uses the project-generated service request (planning goals), the track equipment allocation (initial state), and an addition to adding the detail of antenna sub- antenna operations knowledge base. DPLAN system allocation, the initial schedule under- uses both hierarchical-task network-planning goes continual modification as a result of techniques and operator-based planning changing project needs, equipment availabili- techniques to synthesize these operations ty, and weather considerations. Responding plans. By allowing both operator-based and to changing context and minimizing disrup- hierarchical-task network representations, the tion while rescheduling are key issues. antenna operations knowledge base allows a In 1993, the OMP-26M scheduling system modular, declarative representation of anten- was deployed, partially automating the na operations procedures. In contrast, consid- scheduling of the network of 9-, 11-, and 26- er the two non-AI alternatives proposed: (1) meter (m) antennas. Use of OMP-26M resulted operations script and (2) an exhaustive in a fivefold reduction of scheduling labor, library of plans. Neither operations scripts and network use doubled. The demand-access nor an exhaustive library of plans explicitly network scheduler (DANS) (Chien, Lam, and record the generality and context presumed Vu 1997) is an evolution of the OMP-26M sys- by operations procedures. Planning represen- tem designed to deal with the more complex tations’ explicit representation of such infor- subsystem and priority schemes required to mation should make them easier to maintain schedule the larger 34- and 70-m antennas. as DSN equipment and operations procedures Because of the size and complexity of the evolve. rescheduling task, manual scheduling is pro- DPLAN was initially demonstrated in Febru- hibitively expensive. Automation of these ary 1995 for Voyager downlink, telemetry scheduling functions is projected to save mil- tracks at the DSS-13 antenna at Goldstone, lions of dollars a year in DSN operations California. DPLAN is currently being integrated costs. into the larger network monitor and control DANS uses priority-driven, best-first, con- upgrade being deployed at DSN stations that straint-based search and iterative optimiza- will enable automation projected to reduce tion techniques to perform priority-based DSN operations costs by over $9 million a rescheduling in response to changing net- year. The current DPLAN system supports the work demand. With these techniques, DANS full range of 34-m and 70-m antenna types, first considers the antenna-allocation process all standard service-request classes, and

108 AI MAGAZINE Articles approximately 20 subsystems. Figure 2 shows a plan generated by DPLAN to perform precali- missing lines filled format of image is BYTE bration of a 34-m beam waveguide antenna single pixel spikes removed Number of lines is 800 to provide a telemetry, commanding, and manually navigate images Number of samples is 800 update archival sedr Mission is Galileo ranging services track. Current work on the map project Sensor is SSI DPLAN system focuses on enhancing the sys- construct mosaic Picture Number is GIJ0046 Target is Jupiter tems replanning capability and its ability to Spacecraft Event Timer Year is 1996 represent and reason about plan quality … (Chien, Hill, et al. 1996). Planning Systems for Science Data Preparation and Analysis In the MVP (Chien and Mortensen 1996; Chien 1994) project, planning techniques are being applied to develop a system to automatically generate procedures to satisfy incoming science requests for data. MVP allows a user to input image-processing goals based on the availability and format of relevant image data and produces a plan to achieve the image-processing goals. This plan is then translated into an executable VICAR program. In contrast, manual writing of VICAR scripts is a knowledge-intensive, labor-intensive task requiring knowledge of image-processing techniques, knowledge of image Figure 3. Goals, Initial State, and Raw Images.

compute-om-matrix

OMCOR2

tiepoint-file project mosaic-file-list

MANMATCH

refined-overlap-pairs mosaic-file-list project EDIBIS

crude-overlap-pairs

MOSPLOT-construct-crude

default-nav IBISNAV mosaic-file-list latlon initial predict source CONSTRUCT MOSAIC FILE LIST target GLL_LL effect GLLCAMPAR SHADOWED = operator project underlined = top-level goal operator italics = initial state satisfied condition normal = operator precondition satisfied by effect lines are drawn from operator preconditions to operators to effects precondition

Figure 4. Subplan.

SPRING 1997 109 Articles

ence products for the Galileo Jupiter encounter. For radiometric correction, color- Conceptual Steps VICAR Code triplet reconstruction, and simple mosaicking

get initial navigation IBISNAV OUT="file_list.NAV" PLANET=target_0_10 + tasks, MVP reduces the effort to generate an information PROJECT="GLL " SEDR=@RIMSRC FILENAME="file_list.ilist" initial VICAR script for an expert analyst !! Construct initial overlap pairs MOSPLOT construct initial MOSPLOT inp="file_list.NAV" nl=lines_0_6 ns=samples_0_6 project="GLL " from 4 hours to 15 minutes and for a novice overlap pairs ! mos.overlap is just a holder for the overlap plot. dcl copy printronx.plt mos.overlap analyst from several days to 4 hours. dcl print/nofeed mos.overlap Figure 3 shows the beginning of the succes- refine initial !! Refine initial overlap pairs edibis overlap pairs EDIBIS INP="file_list.OVER" sion of steps. At the top, the image-process- !! Manmatch mosaic file list ing goals (stated by the user using a GUI), !! If there is no existing tiepoint file..... !! Check if a tiepoint file exists. image state (derived automatically from the !! The following code is in written VMS image database), and input raw images are !! LOCAL STR STRING INIT = "" LET _ONFAIL = "CONTINUE" !! Allow the pdf to continue shown. In figure 4, the plan structure for a find previous !! if a file is not found. tiepoint file DCL DEASSIGN NAME portion of the overall plan is shown. In this (if present) DCL DEFINE NAME 'F$SEARCH("file_list.TP") LOCAL STR STRING graph, nodes represent image-processing TRANSLOG NAME STR LET _ONFAIL = "RETURN" !! Set PDF to return on error actions and required image states to achieve IF (STR = "") the larger image-processing goal. Figure 5 MANMATCH INP=("file_list.NAV","file_list.OVER") + use manmatch OUT="file_list.TP" PROJECT="GLL " 'SEDR FILENAME="file_list.ILIST" shows the actual MVP-generated VICAR code, program to !! If an old tiepoint file exists... construct or !! The old tpfile is part of input and later overwritten. highlighting the correspondence to the refine tiepoint ELSE image-processing steps in the plan. Finally, in file MANMATCH INP=("file_list.NAV","file_list.OVER","file_list.TP") + OUT="file_list.TP" PROJECT="GLL " 'SEDR FILENAME="file_list.ILIST" figure 6, we have the produced (output) use tiepoints !! OMCOR2 OMCOR2 INP=("file_list.NAV","file_list.TP") PROJECT="GLL " GROUND=@GOOD science product—a mosaic of Jupiter’s Red to construct OMCOR2 INP=("file_list.NAV","file_list.TP") PROJECT="GLL " GROUND=@GOOD OM matrix Spot constructed from the raw images received from the Galileo spacecraft currently in orbit around Jupiter. Figure 5. MVP-Produced VICAR Code. Although MVP successfully automates certain image-processing tasks, in deploying MVP, we learned the high cost of knowledge base maintenance (approximately 0.8 work-years of effort during the first year of operation). Correspondingly, current work focuses on knowledge base analysis tools to assist in the process of planning knowledge base development, validation, and maintenance (Chien 1996). Machine Learning for Large-Scale Planning and Scheduling Although most scheduling problems are NP- complete in worst-case complexity, in practice, for specific distributions of problems, domain-specific search strategies have been shown to perform in much better than expo- nential time. However, discovering these search strategies is a painstaking, time-consuming process that requires extensive Figure 6. Final Images Produced by MVP. knowledge of both the domain and the scheduler. The goal of adaptive problem solving (APS) is to automate this process of cus- database organization and metadata conven- tomization by learning heuristics specialized tions, and knowledge of the VICAR program- to a distribution of problems. ming environment. Analysts require several Our APS work focuses on statistical learning years to become VICAR experts, and produc- methods used to search for and recognize ing each VICAR script takes hours to months superior problem-solving customizations and of analyst effort. Automating the filling of heuristics (Gratch and Chien 1996; Chien, more routine tasks by the MVP system frees up Gratch, and Burl 1995). Work to date has expert analyst time for more complex, unique achieved strong results. Using the LR-26 processing requests. scheduler on scheduling data for 1996 to MVP is currently being used to produce sci- 1997, statistical machine-learning techniques

110 AI MAGAZINE Articles

0.120

0.100 Expert Strategy 0.080

0.060

0.040 Learned Strategies

0.020 Random Strategies

0 0 30 60 90 120 150 180 210 240 270

Figure 7. Estimated Expected Utility of Machine-Derived, Human Expert–Derived, and Randomly Generated Strategies for Antenna Scheduling. found strategies that improved on human oversimplify the initial task of symptom iden- expert–derived strategies by decreasing central tification. In such work, anomaly detection processing unit time for solvable problems by typically does little more than limit-sensing 50 percent and increasing solvable problems sensor data against static, manually by 15 percent. We are currently extending predefined red lines, or predictions of expen- techniques to allow for specialization of con- sive simulations. However, the smaller bud- trol strategies as directed by empirical learning gets and novel challenges of future NASA methods and allow for control of constraint missions demand cheap, robust, and early relaxation to improve schedule quality. detection to maximize the opportunities for Figure 7 illustrates the effectiveness of the low-cost preventive operation. Because com- machine-learning techniques. It shows the plex NASA domains typically contain both probability density functions for the utility of large volumes of both engineering sensor strategies derived using APS, the human data and human expertise, our collective expert–derived strategy, and randomly select- work at JPL pushes both machine-learning ed strategies in the entire encoded strategy and knowledge engineering methods but space. In this data set, higher utility corre- strives to find an effective balance. In the following paragraphs, we summarize our recent sponds to lower problem-solving time work in these areas. (toward the left of the graph), and the data Initial work in this area focused on the task clearly indicate the superiority of the APS- of continually identifying which sensors are derived strategies over the human currently the most interesting, using informa- expert–derived strategy (which is already sig- tion-theoretic metrics. This work led to the nificantly better than an average strategy in selective monitoring (SELMON) system (Doyle the complete strategy space). 1995; Doyle et al. 1993). SELMON compares histograms of current data against those of Monitoring and Diagnosis historic data, identifying sensors whose binned frequency distributions are more dis- To address the key NASA goals of faster, tinct than historic (expected) ones. It can also cheaper, and better mission operations, the use causal orderings among the sensors to MDT group and the MLS group have been help conditionalize and isolate anomalies. developing a set of complementary methods SELMON overcomes the oversimplicity of tradi- that are suitable for both on-board and tional limit sensing but ignores the sig- ground-based robust anomaly identification. nificance of global context and temporal Both in research and in practice, diagnosis ordering. Thus, it is most appropriate for data work on complex analog domains tends to sets that are statistically stationary.

SPRING 1997 111 Articles

The ELMER (envelope learning and monitor- Automating this process involves not only ing using error relaxation) system, recently detecting and summarizing anomalies but developed by the MDT group (DeCoste 1996), also providing enough detail for ground engi- focuses on the task of automated real-time neers to verify whether the remaining behav- detection of anomalies in time-series data. It ior is truly nominal. Traditional summariza- specifically addresses the issues of context tion techniques focus on fixed gross statistics dependencies and nonstationary time-series (for example, mins, maxs, means), data. It is being field tested in several NASA prespecified event logging, and information- domains, including the space shuttle, the theoretic compression. extreme-ultraviolet explorer, and DSN. Each The goal of ESUM is to automatically select such complex real-world domain requires subsets of downlink data (that is, specific sen- monitoring thousands of sensors and pro- sor values for specific times) sufficient for vides millions of historic samples of each sen- ground operators to perform diagnosis and sor for training. verification tasks. ESUM can also, in part, be ELMER provides a data-driven, iterative, viewed as intelligent prefetching of data that three-stage approach to multivariant time- are likely to be desired if current trends lead series prediction: (1) systematic selection of to clear anomalies later. Thus, it might prove input feature reencodings (for example, time- useful for data archive compression manage- windowed mins, maxs, means, derivatives, ment (both on board and at ground) as well. lag vectors), (2) greedy (linear-time complexi- The basic approach is to select those data that ty) nonlinear feature construction (for exam- ELMER’s trained and adaptive envelope func- ple, products, sigmoidals), and (3) linear tions indicate are most relevant to detecting The ELMER regression (with relaxed error metrics). ELMER’s anomalies and prespecified events. We are system third stage is particularly novel: explicitly currently exploring this work for two future learning two separate functions, for high- and NASA missions: a Pluto express flyby and a focuses low-expectation bounds (envelopes) for beacon experiment on New Millennium Deep on the future data versus one traditional overall least Space One. squared fit. By starting with each envelope as The MLS group is applying machine-learn- task of the two static red-line values traditionally ing techniques to the problem of modeling automated used in monitoring operations, ELMER can engineering time-series data as well. The incrementally tighten the values toward sim- group has developed a prototype software real-time ulation-quality function bounds in an any- package that segments telemetry streams for detection of time manner. individual sensors into distinct and statisti- anomalies in ELMER differs notably from common alter- cally significant modes of activity. These algo- natives, such as neural networks, by being rithms provide a basis for an automated time-series highly constructive and using novel error mechanism for identifying and classifying data. metrics that are particularly appropriate for distinct system modes of operation. The soft- the constraint-checking nature of monitoring ware makes use of probabilistic and hierarchi- tasks (versus prediction as such). In particular, cal segmentation algorithms (such as hidden we can bias the error metrics to avoid false Markov models) that observe past telemetry alarms (which commonly plague other auto- streams and identify distinct regimes of activ- mated approaches) at the expense of obtain- ity based on elemental features of the sensor ing bounds that are weaker than typical neu- time series. The software has been applied ral network predictions (but still much tighter successfully to various space shuttle and than traditional red lines). We are exploring extreme ultraviolet explorer satellite teleme- several extensions to this work, including try streams. Currently, we are investigating using both hidden Markov model learning the application of this technology to future and qualitative reasoning to better guide fea- spacecraft design and operations, with poten- ture construction and selection. tial benefit in the areas of prediction, anoma- We are also extending ELMER to address the ly detection and diagnosis, query by content, more general problem of summarizing the and data summarization and compression. behavior of sensor data over large windows of Whereas the previous work focused on time, in a project called engineering data sum- data-driven time-series prediction and ma- marization (ESUM). For example, a planned chine-learning techniques, it is also impor- mission to Pluto would spend nearly eight tant to cost-effectively leverage the large base years in mostly uneventful cruise mode, for of human expertise available in NASA which automated weekly summaries of behav- domains—particularly because historic sensor ior (with low-downlink bandwidth) would be data for complex systems such as spacecraft critical to achieving low operations costs. are seldom fully representative of future

112 AI MAGAZINE Articles behaviors. Thus, the MDT group has also Knowledge Discovery been developing methodologies and semiau- tomated tools for knowledge engineering, and Data Mining for integrating, and testing of qualitative and Scientific Data Analysis quantitative spacecraft models suitable for The MLS group is heavily involved in a num- robust monitoring and diagnosis. ber of projects designed to understand and Our current work in modeling and testing exploit large-scale scientific data sets. These focuses on the Deep Space One Mission. We efforts can roughly be classified into two are streamlining the process of designing and groups: One group consists of data-mining implementing on-board monitoring software methods focused on the extraction of scien- with the creation of a family of reusable mon- tific insight from massive data sets, usually itoring components that are easily configured on the order of gigabytes to terabytes in size. to detect important changes in status. The These data are typically collected by space- resulting status information feeds into the borne and other sensors and are then ar- remote-agent mode-identification module chived and managed on the ground. This (Williams and Nayak 1996) for mode tracking work features a mixture of techniques drawn and diagnosis. Our monitoring components from machine learning, statistics, databases, perform functions such as noise and transient and high-performance computing. The other Deep Space filtering, min-max thresholding, sensor-limit major theme is the transference of machine detection, phase-plane analysis, and summa- learning and data-mining successes to the One also rization for a much-reduced telemetry stream. realm of space-borne computing and autono- challenges In the iterative spiral model of spacecraft my. By performing intelligent data processing design and development used for Deep Space on board spacecraft, new scientific ex- conventional One, software and hardware modules need to periments are enabled that exploit and verification be integrated several times through the pro- enhance the capabilities of autonomous ject. By reusing metainformation already avail- spacecraft. These science-driven autonomy and able for defining software interfaces as mes- ideas complement other work being conduct- validation sages and function calls, we are exploring the ed in the AI and MDT groups, with the com- because its use of graph-dependency analysis techniques mon goal of enabling efficient autonomous (Rouquette 1996) for coordinating the integra- spacecraft. fast-paced tion of tightly coupled software tasks, tracking The data-mining work covers a broad spec- spiral- complex multitask software interactions, and trum, including nearly every major scientific detecting behavioral discrepancies at the level subdiscipline relevant to NASA. These subdis- development of message passing and function calls. ciplines include atmospheric and earth sci- process Deep Space One also challenges convention- ence, planetary science, and solar physics. al verification and validation because its fast- Several prominent themes are shared by each does not paced spiral-development process does not of these projects: an emphasis on data-driven generate generate formal testable requirements. modeling, the application of computationally Instead, testing is driven by mission scenar- intensive statistics as a major tool, and the formal ios, and test behavior is filtered through flight use of a number of prominent machine-learn- testable rules and localized episodes of expected ing techniques such as decision trees and requirements. events. A perfect test is one that satisfies posi- neural networks. At the same time, there are tive flight rules, avoids negative flight rules, distinct differences: Some are more crucially and accounts for every data log as part of an dependent on high-performance computing expected episode. Our objective here is to cre- than others (although in the long run, all will ate a scenario-based test methodology (and be driven in this direction as NASA data sets tools) that finds the bugs without requiring dramatically increase in size). Several are hard-to-obtain requirement specifications. deeply concerned with temporal and spa- Flight software in modern spacecraft com- tiotemporal processes, but others focus on bines event-driven (reactive) programming the accurate analysis of spatial patterns only. with continuous control. Typically, both have A major challenge for the future is to develop been implemented in a procedural language a systematic approach to this range of prob- such as C, but reactive designs beg for a situa- lems that allows powerful algorithms to be tion-action directive in the language so that applied in many contexts but also accounts programmers can specify what to do rather for the subtleties of individual data sets and than when and where to do it. Toward this problems in the process. goal, we are exploring the use of R++, a tight As an example of the scientific insight that integration of C++ and path-based rules can be obtained by merging data-mining (Crawford et al. 1996). ideas with scalable computation, consider the

SPRING 1997 113 Articles

Figure 8. Ground-Displacement Map for Landers Earthquake Region Generated Using Statistical Learning Techniques. Direction of ground displacement is indicated in grey-scale wheel at left.

problem of automatically detecting and cata- important scientiﬁc information from even loging important temporal processes in mas- high-resolution images. This process is time sive data sets. In general, this task is one of consuming and extremely expensive. In a col- overwhelming scale that has so far eluded laboration between the MLS group and the automation for the great majority of scientiﬁc Terrestrial Science Element at JPL, the problems. Historically, careful manual inspec- QUAKEFINDER system was developed and imple- tion of images by highly trained scientists has mented as a prototype data-mining system been the standard method of extracting that dramatically speeds up this process

114 AI MAGAZINE Articles

(Stolorz and Dean 1996). It tackles the prob- sky objects using the machine-learning com- lem of analyzing the earth’s crustal dynamics ponent of our system. The JPL learning algo- by enabling the automatic detection and rithms GID3*, O-B TREE, and RULER have been measurement of earthquake faults from satel- used to produce decision trees and lite imagery. The system has been used to classification rules from training data consist- map the direction and magnitude of ground ing of astronomer-classified sky objects. These displacements that resulted from the 1992 classifiers are applied to new survey images to Landers earthquake in southern California obtain the classifications needed for a com- over a spatial region of several hundred plete northern celestial hemisphere survey square kilometers at a resolution of 10 m to a database containing on the order of 5 x 107 (subpixel) precision of 1 m. Figure 8 shows galaxies, 2 x 109 stars, and more than 105 this ground-displacement map. The grey-scale quasars. wheel at the bottom left indicates the direc- One notable success of the SKICAT system is tion of inferred ground movement. The fault its use by Caltech astronomers, with color itself is clearly shown by the discontinuity in and classification information, to select the ground movement. This calculation is the quasar candidates at red shifts of z > 4 first to have been able to extract area-mapped (Kennefick et al. 1995). This approach was 40 information about two-dimensional tectonic times more efficient in terms of the number processes at this level of detail. It is accom- of new quasars discovered for each unit of plished using a combination of statistical telescope time than the previous survey for inference (entropy minimization–style learn- such objects done at Palomar. To date, some The volume ing), parallel computing, and global opti- 24 new quasars at these high red shifts have of image data mization techniques. been found and have been used to constrain Although applied initially to the specific the evolution of their luminosity function, collected by problem of earthquake fault analysis, the indicating the epoch of quasar and, probably, spacecraft principals used by QUAKEFINDER are broadly galaxy formation. They have also been used applicable to a far more general class of calcu- by many other groups of astronomers to has reached lations involving subtle change detection in probe the early universe and the early inter- a level that high-resolution image streams. Success in this galactic medium. makes the domain points the way to a number of such SKICAT is now being extended to perform data-mining calculations that can directly unsupervised clustering on the roughly 40- traditional and automatically measure important tempo- dimensional feature space produced by the approach of ral processes to high precision from massive initial image-processing software. Probabilis- data sets such as problems involving global tic clustering methods are being applied to manually climate change and natural hazard monitor- systematically explore the parameter space of examining ing as well as general image-understanding object measurements in an unbiased fashion tasks involving target detection and iden- to search for possible previously unknown each collected tification in noisy image streams. Efforts are rare groupings. It is quite possible that in a image now under way to use QUAKEFINDER to analyze data set this large, some previously unno- infeasible. possible activity on the ice-covered surface of ticed, new astronomical type of objects or Jupiter’s moon Europa and search for sand phenomena can be found by suitable outlier dune activity on Mars. There are also a num- searches. This result would be major and path ber of potential applications in the biomedi- breaking, both displaying scientific interest cal imaging field. and demonstrating the power of machine- In a collaborative effort with Caltech assisted discovery in astronomy. astronomers, machine-learning techniques Although much of the focus of the work at have also been applied to a problem in the JPL is focused outward toward the planets and area of large-image database analysis. SKICAT beyond, there are many important unan- (Weir, Fayyad, and Djorgovski 1995; Fayyad, swered scientific questions concerning the Djorgovski, and Weir 1996a, 1996b) inte- earth’s geology and climate, which are also grates image processing, high-performance under investigation. The MLS group is devel- database management, and AI classification oping novel spatiotemporal data-mining tech- techniques to automate the reduction of the niques and tools for tapping the vast resources second Palomar Observatory sky-survey of earth-observed data. One example is our image database. Image-processing routines ongoing analysis of low-frequency variability first detect sky objects and measure a set of in the upper atmosphere, specifically spatial important features for each object, for exam- grids of 700-megabyte geopotential height ple, brightness, area, extent, and mor- records taken daily since 1947 over the North- phology. These features are used to classify ern Hemisphere (available from the National

SPRING 1997 115 Articles

Figure 9. Automatically Classified Solar Image. Top Left: Raw solar image. Top Right: Threshold-derived classifications. Bottom Left: Markov random field–derived classification.

probabilistic method for determining the best number of clusters (Smyth 1996), we have provided the first objective proof of the existence of three distinct regimes in the earth’s upper atmosphere. This result is significant from both a scientific and a methodological viewpoint: It has answered a long-standing open question in the field and is the first application of objective cluster-validation criteria in this area. Ongoing work is focusing on temporal clustering using generalized hidden Markov models and extensions to oceanic data and ocean-atmosphere modeling. The MLS group is also developing statistical pattern- recognition methods leading to automatic and objective image-analysis systems for science data sets in astronomy and solar physics (Turmon and Pap 1997). The system allows scientists to label active regions on solar images and combine this information with domain-specific knowledge to train a pattern recognizer. Data sources include a 30-year database of ultraviolet intensity images taken each day on the ground and light images and mag- Oceanic and Atmospheric netic field maps taken many times daily from the Administration). A key scientific NASA–European Space Agency SOHO (solar and helio- question is whether recurrent, spheric observatory) satellite. Our existing recognition stable regimes of climatic activi- system incorporates information about pixel-level spatial ty exist. We applied finite-mix- continuity. Ongoing efforts will allow higher-level con- ture models to the data to see if structs describing active regions; also, we want to make an we could discover the underly- active-region database and correlate it with existing solar- ing cluster structure. Combining irradiance data for climatological purposes. Figure 9 illus- the cluster results in a novel trates the data classification enabled by the statistical pat-

116 AI MAGAZINE Articles

Figure 10. Volcanoes on Venus. Top: Volcano-appearance components (in decreasing strength). Right: A 75-kilometer x 75-kilometer region of Venus containing numerous small volcanoes labeled by experts.

tern-recognition methods. At the top left is the input solar data; the top right image is a classification produced by a simple thresholding algorithm. At the bottom left is the (improved) classification produced using the Markov random-field pattern-recognition techniques. The volume of image data collected by spacecraft has reached a level that makes the traditional approach of manually examining each collected image infeasible. The scientists, who are the end users of the data, can no longer perform global or comprehensive analyses effectively. To aid the scientists, we have developed a trainable pattern-recognition system, known as JARTOOL (JPL adaptive recognition tool) (Fayyad et al. 1995; Burl et al. 1994), for locating features of interest in remote sensing data. The system has initially been applied to the problem of locating small volcanoes in Magellan synthetic aperture radar imagery of Venus. Scientists train the system by labeling volcano examples with a simple GUI. The system then learns an appearance model for volcanoes and uses this space missions. As spacecraft become highly model to find other instances of volcanoes in autonomous, there are increasing opportuni- the database. Figure 10 shows the radar-input image data at the left and volcano-appear- ties for in situ analysis of scientific data, ance components at the right (in decreasing enabling real-time planning of scientific strength). It is particularly interesting to note experiments in response to important events. that the first several appearance components In collaboration with planetary scientists at correspond to descriptive features used by the Southwest Research Institute in Boulder, human experts, such as a pit at the summit Colorado, a prototype system has been devel- and a shadow opposite the look direction of oped that performs automatic on-board the radar. detection of natural satellites in the vicinity The MLS group is also applying AI and pat- of asteroids and planets (Stolorz et al. 1997). tern-recognition techniques to the problem The initial system deals with the case of of conducting on-board science analysis for static objects and is now being extended to

SPRING 1997 117 Articles

which is designed to provide high-performance low-power multicomputers for space- borne platforms. The intent is to provide a scalable family of computing platforms for a wide range of spacecraft. It is anticipated that autonomy applications developed by the MLS group will be used by science instrument principal investigators on future space missions. These autonomy applications will then, in turn, be major users of this future space-borne scalable computing capability.

Design Automation Spacecraft design is a knowledge- and labor- intensive task with demanding requirements. Spacecraft must survive the most hostile envi- ronments, operating at great distances and Figure 11. Automatic Satellite Detection. with little interaction with earth-based per- Top: Raw-data input for infrared filter. Middle: Raw-data input for green filter. sonnel. In synthesizing a design, engineers Bottom: Processed detection image for static satellite detection. must balance rigorous survivability and science requirements against mass, power, volume, and processor limits, all within reduced account for parallax effects that result from mission budgets. To assist spacecraft design both spacecraft and satellite motion. engineers in this challenging task, the AI Accounting for parallax effects involves the group has been developing and deploying introduction of predictive methods in the technology in the area of intelligent opti- presence of uncertainty as well as careful reg- mization. istration techniques. Images from a prelimi- Spacecraft design optimization is difficult nary demonstration of the static satellite- using current optimization methods because detection capability are shown in figure 11. (1) current methods require a significant The top and middle images show two of the amount of manual customization by the four raw-data input images (each taken with a users to be successful and (2) traditional different filter). The bottom image is the pro- methods are not well suited for mixed dis- cessed image used for satellite detection using crete-continuous, nonsmooth, and possibly the prediction and registration processing. probabilistic cost surfaces that can arise in Another system deals with the automatic many design-optimization problems. Of par- analysis of ultraviolet spectra to allow deci- ticular interest are the so-called black-box sion making about the optimal way to con- optimization problems in which the structure duct a spectral experiment. By automatically of the cost function is not directly accessible identifying and removing the main chemical (that is, the cost function is computed using a species present in a scientific target, such as a simulation). cometary tail, decisions can be made about We are currently developing the optimiza- the optimal data-taking mode, for example, tion assistant (OASIS) (Fukunaga, Chien, et al. whether to scan the spectrometer across por- 1997), a tool for automated spacecraft design tions of the tail in a survey mode or to con- optimization that addresses these two issues. centrate on obtaining high-resolution data The goal of OASIS is to facilitate rapid what-if from one area where unexpected species analysis of spacecraft design by developing a might have been discovered. generic spacecraft design-optimization system We believe that in situ machine-learning that maximizes the automation of the opti- applications such as these will have a dramat- mization process and minimizes the amount ic effect on the range and quality of science of customization required by the user. that can be pursued by spacecraft missions in OASIS consists of an integrated suite of glob- the future. They have immediate relevance to al optimization algorithms (including genetic NASA initiatives such as JPL’s New Millenni- algorithms, simulated annealing, and auto- um Program. They also tie in strongly with mated response-surface methods) that are longer-term research and development pro- applicable to difficult black-box optimization jects. One of these projects is JPL’s Remote problems and an integrated intelligent agent Exploration and Experimentation Program, that decides how to apply these algorithms to

118 AI MAGAZINE Articles a particular problem. Given a particular spacecraft design-optimization problem, OASIS performs a metalevel optimization to (1) 0.5 select an appropriate optimization technique to apply to the problem and (2) automatically 0 adapt (customize) the technique to fit the problem. This metalevel optimization is guid- -0.5 ed by both domain-independent and domain-specific heuristics that are automati- -1 cally acquired through machine-learning techniques applied to a database of perfor- 3.5 3 mance profiles collected from past optimiza- 0.5 2.5 Outside2 Diameter (inches) tion episodes on similar problems. 1 1.5 1.5 1 OASIS Total Length (feet) We have been applying the system to 0.5 the problem of penetrator design. A penetrator is a small, robust probe designed to impact a surface at extremely high velocity with the goal of performing sample analysis. Specifically, we have been applying OASIS to design and simulation data from the New 0.5 Millennium Deep Space Two Mission penetrator design in which the design variables are 0 penetrator diameter and length. The Deep Space Two Mission consists of a pair of pene- -0.5 trators to be launched in 1998, impacting the planet Mars to perform soil analysis in 1999. -1

Figure 12 shows optimization surfaces derived 3.5 from impact simulations for candidate 3 0.5 2.5 designs. Each of the three surfaces represents Outside2 Diameter (inches) 1 1.5 the predicted depth of penetration for a dif- 1.5 1 Total Length (feet) ferent soil consistency; in all three cases, the 0.5 surface is highly discontinuous. The negative values represent physically unrealizable designs, and the zero values indicate cases in which the penetrator deﬂects on impact 0.5 (catastrophic mission failure). The goal of the design problem is to produce a physically 0 realizable design while it maximizes the chance of successful penetration over an expected distribution of soil consistencies -0.5 and minimizes design cost -1

3.5 Summary 3 0.5 2.5 This article has described ongoing AI activi- Outside2 Diameter (inches) 1 1.5 1.5 1 ties at JPL. Because of space, time, and coordi- Total Length (feet) nation constraints, we were unable to fully 0.5 cover all related areas of work at JPL (most notably, this article does not cover consider- Figure 12. Optimization Surfaces for Impact Penetration Depth (Output) as able robotics, computer vision, neural net, a Function of Outside Diameter and Length of New Millennium Deep Space fuzzy logic, and pattern-recognition work). Two Mission Penetrator (Three Different Soil Densities). For further information on the projects described in this article, readers are invited to visit our web page at www-aig.jpl.nasa.gov/ or the JPL general web page at www.jpl.nasa. gov// or to contact one of the following indi- viduals at ﬁrstname.lastname @jpl.nasa.gov: Chester Borden, Operations Research Group; Dennis DeCoste, Monitoring and Diagnosis

SPRING 1997 119 Articles

Technology Group; Steve Chien, Artificial Chien, S.; Govindjee, A.; Estlin, T.; Wang, X.; and Intelligence Group; Sven Grenander, Hill Jr., R. 1996. Integrating Hierarchical Task Net- Sequencing Automation Research Group; and work and Operator-Based Planning Techniques to Paul Stolorz, Machine Learning Systems Automate Operations of Communications Anten- nas. IEEE Expert 11(6): 9–11. Group. Chien, S.; Govindjee, A.; Estlin, T.; Wang, X.; Acknowledgments Griesel, A.; and Hill Jr., R. 1997. Automated Genera- tion of Tracking Plans for a Network of Communi- This article describes work conducted by the cations Antennas. In Proceedings of the 1997 IEEE Jet Propulsion Laboratory, California Institute Aerospace Conference, 343–359. Washington, D.C.: of Technology, under contract with the IEEE Computer Society. National Aeronautics and Space Administra- Chien, S. A.; Hill, Jr., R. W.; Govindjee, A.; Wang, tion (NASA). The contributors to this article X.; Estlin, T.; Griesel, M. A.; Lam, R.; and Fayyad, K. want to acknowledge the long-standing sup- V. 1996. Robust, Multi-Layered Plan Execution and port of Melvin Montemerlo of NASA head- Revision for a Network of Communications Anten- quarters, who has fostered the Artificial Intel- nas. Paper presented at the 1996 AAAI Fall Sympo- ligence Program, now the Autonomy sium on Plan Execution, 9–11 November, Cam- Technology Program, for 10 years and has bridge, Massachusetts. personally managed the people and research Chien, S.; Hill Jr., R.; Wang, S.; Estlin, T.; Fayyad, K.; investments that now set the stage for truly and Mortensen, H. 1996. Why Real-World Planning meaningful contributions to NASA goals on Is Difficult: A Tale of Two Applications. In New Directions in AI Planning, eds M. Ghallab and A. the part of AI researchers and technologists at Milani, 287–298. Amsterdam: IOS Press. the Jet Propulsion Laboratory (JPL) and else- Crawford, J.; Dvorak, D.; Litman, D.; Mishra, A.; where. The deep space network (DSN) ele- and Patel-Schneider, P. F. 1996. Path-Based Rules in ments of this work were funded by the DSN Object-Oriented Programming. In Proceedings of Technology Program, managed by Chad the Thirteenth National Conference on Artificial Edwards of the Telecommunications and Mis- Intelligence, 490–497. Menlo Park, Calif.: American sion Operations Directorate of JPL. Association for Artificial Intelligence. DeCoste, D. 1996. Learning and Monitoring with References Families of High-Low Envelope Functions: Maxi- Burl, M.C.; Fayyad, U.M.; Perona, P.; Smyth, P.; and mizing Predictive Precision While Minimizing False Burl, M. P. 1994. Automating the Hunt for Volca- Alarms, Technical Document D-13418, Jet Propul- noes on Venus. In Proceedings of the IEEE Comput- sion Laboratory, Pasadena, California. er Society Conference on Computer Vision and Pat- Doyle, R. J. 1995. Determining the Loci of Anoma- tern Recognition, 302–309. Washington, D.C.: lies Using Minimal Causal Models. In Proceedings Institute of Electrical and Electronics Engineers. of the Fourteenth International Joint Conference Chien, S. 1994. Using AI Planning Techniques to on Artificial Intelligence, 1821–1827. Menlo Park, Automatically Generate Image-Processing Proce- Calif.: International Joint Conferences on Artificial dures: A Preliminary Report. In Proceedings of the Intelligence. Second International Conference on Artificial Intelli- Doyle, R. J.; Chien, S. A.; Fayyad, U. M.; and Wyatt, gence Planning Systems, 219–224. Menlo Park, Calif.: E. J. 1993. Focused Real-Time Systems Monitoring AAAI Press. Based on Multiple Anomaly Models. Paper present- Chien, S. A. 1996. Intelligent Tools for Planning ed at the International Workshop on Qualitative Knowledge Base Development and Verification. In Reasoning about Physical Systems, May, Eastsound, Advances in Knowledge Acquisition, eds. N. Shadbolt, Washington. K. O’Hara, and G. Schreiber, 321–337. Lecture Doyle, R.; Gor, V.; Man, G.; Stolorz, P.; Chapman, Notes in Computer Science. Berlin: Springer Verlag. C.; Merline, W.; and Stern, A. 1997. Autonomy Chien, S. A., and Mortensen, H. B. 1996. Automat- Enables New Science Missions. In Proceedings of ing Image Processing for Scientific Data Analysis of the First Conference on Future Space Science and a Large Image Database. IEEE Transactions on Pat- Earth Science Missions, Space Technology and tern Analysis and Machine Intelligence 18(8): Applications International Forum. Albuquerque, 854–859. N.M.: Institute for Space and Nuclear Power Stud- Chien, S. A.; Gratch, J. M.; and Burl, M. C. 1995. ies. On the Efficient Allocation of Resources for Fayyad, U. M.; Djorgovski, S. G.; and Weir, N. Hypothesis Evaluation: A Statistical Approach. IEEE 1996a. Automating the Analysis and Cataloging of Transactions on Pattern Analysis and Machine Intelli- Sky Surveys. In Advances in Knowledge Discovery in gence 17(7): 652–665. Databases, eds. U. Fayyad, G. Piatetsky-Shapiro, P. Chien, A.; Lam, R.; and Vu, A. 1997. Resource Smyth, and R. Uthurusamy, 71–493. Cambridge, Scheduling for a Network of Communications Mass.: MIT Press. Antennas. In Proceedings of the 1997 IEEE Fayyad, U. M.; Djorgovski, S. G.; and Weir, N. Aerospace Conference, 361–373. Washington, D.C.: 1996b. From Digitized Images to Online Catalogs: IEEE Computer Society. Data Mining a Sky Survey. AI Magazine 17(2): 1–66.

120 AI MAGAZINE Articles

Fayyad, U. M.; Smyth, P.; Burl, M. C.; and Perona, P. Based Planning and Scheduling for Shuttle Payload 1995. A Learning Approach to Object Recognition: Operations. In Proceedings of the 1997 IEEE Applications in Science Image Analysis. In Early Aerospace Conference, 325–341. Washington, D.C.: Visual Learning. Oxford, U.K.: Oxford University IEEE Computer Society. Press. Rouquette, N. 1996. Building Steady-State Simula- Fox, G., and Borden, C. 1994. Low Earth Orbiter tors via Hierarchical Feedback Decomposition. In Resource Allocation and Capacity Planning for the Proceedings of the 1996 National Conference on Deep Space Network Using LEO4CAST. Telecommu- Artificial Intelligence, 991–996. Menlo Park, Calif.: nications and Data Acquisition 42(118): 169–178. American Association for Artificial Intelligence. Fukunaga, A.; Rabideau, G.; Chien, S.; Yan, D. Smyth, P. 1996. Clustering Using Monte Carlo 1997. Toward an Application Framework for Auto- Cross-Validation. In Proceedings of the Second mated Planning and Scheduling. In Proceedings of International Conference on Knowledge Discovery the IEEE Aerospace Conference, 375–386. Washing- and Data Mining, 126–133. Menlo Park, Calif.: ton, D.C.: IEEE Computer Society. AAAI Press. Fukunaga, A.; Chien, S.; Mutz, D.; Sherwood, R.; Stolorz, P., and Dean, C. 1996. QUAKEFINDER: A Scal- and Stechert, A. 1997. Automating the Process of able Data-Mining System for Detecting Earthquakes Optimization in Spacecraft Design. In Proceedings from Space. In Proceedings of the Second Interna- of the IEEE Aerospace Conference, 411–427. Wash- tional Conference on Knowledge Discovery and ington, D.C.: IEEE Computer Society. Data Mining, 208–213. Menlo Park, Calif.: AAAI Press. Gat, E. 1997. ESL: A Language for Supporting Robust Plan Execution in Embedded Autonomous Agents. Stolorz, P.; Doyle, R.; Gor, V.; Chapman, C.; Merline In Proceedings of the IEEE Aerospace Conference, W.; and Stern, A. 1997. New Directions in Science- 319–324. Washington, D.C.: IEEE Computer Soci- Enabling Autonomy. In Proceedings of the 1997 ety. IEEE Aerospace Conference, 387–399. Washington, Gratch, J. M., and Chien, S. A. 1996. Adaptive Prob- D.C.: IEEE Computer Society. lem Solving for Large-Scale Scheduling Problems: A Turmon, M. J., and Pap, J. M. 1997. Segmenting Case Study. Journal of Artificial Intelligence Research Chromospheric Images with Markov Random 4:65–396. Fields. In Proceedings of the Second Conference on Statistical Challenges in Modern Astronomy. New Hill Jr., R. W.; Chien, S. A.; Fayyad, K. V.; Smyth, C.; York: Springer. Forthcoming. Santos, T.; and Bevan, R. 1995. Sequence-of- Events–Driven Automation of the Deep Space Net- Weir, N.; Fayyad, U. M.; and Djorgovski, S. G. 1995. work. Telecommunications and Data Acquisition Automated Star-Galaxy Classification for Digitized 42(124). POSS-II. The Astronomical Journal 109(6): 2401–2412. Kennefick, J. D.; De Carvalh, R. R.; Djorgovski, S. Williams, B., and Nayak, P. 1996. Immobile Robots: G.; Wilber, M. M.; Dickinson, E. S.; Weir, N.; AI in the New Millennium. AI Magazine 17(3): Fayyad, U.; and Roden, J. 1995. The Discovery of 17–35. Five Quasars at z > 4 Using the Second Palomar Sky Survey. Astronomical Journal 110(1): 78–86. Loyola, S. J. 1993. PC4CAST—A Tool for Deep Space Network Load Forecasting and Capacity Planning. Telecommunications and Data Acquisition 42(114): 170–194. Muscettola, N.; Smith, B.; Chien, S.; Fry, C.; Rajan, K.; Mohan, S.; Rabideau, G.; and Yan, S. 1997. On- Board Planning for the New Millennium Deep Space Steve Chien is technical group One Spacecraft. In Proceedings of the 1997 IEEE supervisor of the Artificial Intelli- Aerospace Conference, 303–318. Washington, D.C.: gence Group, Information and IEEE Computer Society. Computing Technologies Re- search Section, at the Jet Propul- Pell, B.; Bernard, D.; Chien, S.; Gat, E.; Muscettola, sion Laboratory (JPL), California N.; Nayak, P.; and Williams, B. 1996. Remote-Agent Institute of Technology, where he Prototype for an Autonomous Spacecraft. In Pro- leads efforts in automated plan- ceedings of the SPIE Conference on Optical Sci- ning and scheduling. Chien is ence, Engineering, and Instrumentation, Volume also an adjunct assistant professor with the Depart- on Space Sciencecraft Control and Tracking in the ment of Computer Science at the University of New Millennium. Bellingham, Wash.: Society of Southern California. He holds a B.S., an M.S., and a Professional Image Engineers. Ph.D. in computer science, all from the University Pell, B.; Gat, E.; Keesing, R.; Muscettola, N.; and of Illinois. He has been an organizer of numerous Smith, B. 1996. Plan Execution for Autonomous workshops and symposia, including the 1989 and Spacecraft. Paper presented at the AAAI Fall Sympo- 1991 machine-learning workshops and the Ameri- sium on Plan Execution, 9–11 November, Cam- can Association for Artificial Intelligence (AAAI) bridge, Massachusetts. symposia in 1992 and 1994. In 1995, Chien Rabideau, G.; Chien, S.; Stone, P.; Willis, J.; Egge- received the Lew Allen Award for Excellence. He meyer, C.; and Mann, T. 1997. Interactive, Repair- holds several technical lead positions on JPL and

SPRING 1997 121 Articles

National Aeronautics and Space Administration reasoning and machine learning. He has authored planning and scheduling projects, including colead over 40 publications, with emphasis on automated for the on-board planning engine for the New Mil- modeling and applications in real-time anomaly lennium Deep Space One Mission, lead for the detection in time-series data. He is the developer of automated planning and scheduling engine for the the SELMON system for selective monitoring of New Millennium Earth Orbiter One Mission, lead spacecraft telemetry, with applications at JPL, at for the demand-access scheduling element of the other National Aeronautics and Space Administra- New Millennium Deep Space One Beacon Monitor tion (NASA) centers, and outside NASA. Experiment, and lead for the Automated Planning and Scheduling elements of the Network Automa- tion Work Area of the Deep Space Network Tech- nology Program. Chien has presented invited semi- Paul E. Stolorz is technical group nars on machine learning, planning, and expert supervisor of the Machine Learn- systems; is a AAAI and an International Joint Con- ing Systems Group, Information ference on Artificial Intelligence tutorial presenter and Computing Technologies on automated planning; and has authored numer- Research Section, at the California ous publications in these areas. His current research Institute of Technology (Caltech), interests lie in planning and scheduling, machine Jet Propulsion Laboratory (JPL). learning, operations research, and decision theory. His current research interests are in machine learning, statistical pattern recognition, computational biology and Dennis DeCoste is the technical biochemistry, parallel data mining, and knowledge group leader of the Monitoring discovery. He holds a Ph.D. in theoretical physics and Diagnosis Technology Group from Caltech and a DIC from Imperial College, Uni- in the Information and Comput- versity of London. He currently serves as principal ing Technologies Research Section investigator for several data-mining tasks funded by at the Jet Propulsion Laboratory the National Aeronautics and Space Administration (JPL), California Institute of Tech- at JPL, including the science data-analysis and visu- nology. He received his Ph.D. in alization task, on-board science-analysis task, and computer science in 1994 from the temporal data-mining task. He has also spear- the University of Illinois at Urbana-Champaign in headed the QUAKEFINDER Project at JPL, involving the the area of qualitative physics. Since joining JPL, he detection and measurement of the earth’s crustal has been leading several research and development dynamics on parallel supercomputers. He has been projects on automated monitoring and diagnosis funded by the National Institutes of Health to across a wide variety of National Aeronautics and develop protein secondary-structure prediction Space Administration domains, including space techniques and heads the HIV secondary-structure shuttle, extreme ultraviolet explorer, Pluto express, prediction collaboration headquartered at JPL. This and deep space network. He has published several collaboration has produced RNA secondary-struc- peer-reviewed articles in the areas of automated ture predictions for the HIV genome, the largest monitoring and diagnosis and regularly serves as RNA molecules that have ever been folded by com- reviewer for both Artificial Intelligence and the puter, using the Touchstone Delta supercomputer at American Association for Artificial Intelligence. Caltech. Stolorz has been involved in scientific DeCoste’s current research on automated monitor- applications of machine learning for more than 12 ing and diagnosis focuses on the areas of adaptive years and in massively parallel computing for more time-series prediction, visualization, and clustering; than 15 years, with numerous archival publications stepwise linear and nonlinear regression; automat- in the scientific and machine-learning literature. He ed feature construction; constraint satisfaction; and regularly lectures and conducts tutorials at the Sum- model-based reasoning. mer School on Complex Systems, International Statistics Institute. He has served on the program committee for the Second International Conference Richard Doyle is technical sec- on Knowledge Discovery and Data Mining (KDD- tion manager of the Information 96) and serves as publicity chair and on the confer- and Computing Technologies ence and program committees for KDD-97. Research Section and program manager for the Autonomy Tech- nology Program at the California Institute of Technology, Jet Propulsion Laboratory (JPL). He received his Ph.D. in computer science at the Massachusetts Institute of Technolo- gy Artificial Intelligence Laboratory in 1988. He is U.S. program chair for the Fourth International Symposium on Artificial Intelligence, Robotics, and Automation for Space, to be held in Tokyo in 1997. His research interests are in causal and model-based

122 AI MAGAZINE