An Intelligent Training System for Flight Controllers Space Shuttle

Home , Flight controller

R. Bowen Loftin Lui Wang and Paul Baffes University of Houston-Downtown Artificial Intelligence Section, FM72 One Main Street NASA/Johnson Space Center Houston, TX 77002 Houston, TX 77058

Grace Hua Computer Sciences Corp. 16511 Space Center Blvd. Houston, TX 77058

Abstract study of flight rules, training manuals, An autonomous intelligent training system and on-the-job training (OJT) in inte- which integrates expert system technology grated simulations. From two to four years with training/teaching methodologies is is normally required for a trainee FDO to described. The system was designed to train be certified for many of the tasks for Mission Control Center Flight Dynamics Offi- which he or she is responsible during cers at NASA/JohnsonSpace Center to deploy a Space Shuttle missions. OJT is highly labor certain type of satellite from the Space Shuttle. intensive and presupposes the availability The purpose of the system is to provide a of experienced personnel with both the bridge between the fundamental knowledge time and ability to train novices. As the acquired from textbooks and the procedural number of experienced FDOs has been skills necessary for successful performance in reduced through retirement, transfer large-scale integrated simulations. Trainees (especially of Air Force personnel), and are exposed to scenarios that evolve to higher promotion and as the preparation for and levels of difficulty and are provided with actual control of missions occupies most of assistance tailored to their demonstrated skill the MCC’s available schedule, OJT has level. Such training devices can serve to aug- become increasingly difficult to deliver to ment on-the-job training, insure uniformity in novice FDOs. As a supplement to the training, and allow scarce training expertise to existing modes of training, the Artificial be delivered to a large number of trainees. The Intelligence Section (AIS) at NASA/JSC has architecture of the system is modular and can developed an autonomous intelligent be adapted to a wide variety of training tasks. computer-aided training system. The system trains inexperienced flight con- Problem Definition trollers in the deployment of a Payload- The Mission Operations Directorate Assist Module (PAM) satellite from the Space Shuttle. This task is complex, (MOD) at NASA/Johnson Space Center (JSC) mission-critical, and requires skills used is responsible for the ground control of all by the experienced FDO in performing Space Shuttle operations. Those opera- many of the other operations which are tions which involve alterations in the his or her responsibility. characteristics of the Space Shuttle’s orbit are guided by a flight controller, known as a Flight Dynamics Officer (FDO), sitting Description of the Application at a console in the front room of the Mis- Since the first proposals to apply arti- sion Control Center (MCC). Currently, the ficial intelligence to the tutoring or training of the FDOs in flight operations is training task (Carbonell,1970; Hartley and principally accomplished through the Sleeman, 1973), a large number of systems

105 have been developed for both academic the degree possible, on a single screen, tutoring environments and indus- the environment of the domain in which trial/government training environments the trainee will eventually work. Figure 1 (Sleeman and Brown, 1982; Yazdani, 1986; shows a typical screen display. Kearsley, 1987; Wenger, 1987). In spite of these efforts, few completed systems have Domain Expert come into widespread use for routine A domain expert, referred to as the training. Deploy Expert (DeplEx) is in the form of The training system is designed to aid production rule system that is capable of novice FDOs in acquiring the experience carrying out the satellite deployment pro- necessary to carry out a PAMdeploy in an cess using the same information that is integrated simulation. It is intended to available to the trainee. DeplEx also con- permit extensive practice with both tains a list of "mal-rules" (explicitly iden- nominal deploy exercises and others con- tified errors that novice trainees com- taining typical problems. After success- monly make; Sleeman, 1982) so that the fully completing training exercises which trainee can be provided with feedback contain the most difficult problems specifically designed to help him or her together with realistic time constraints overcome any anticipated conceptual or and distractions, the trainee should be procedural problems. able to successfully complete an integrated simulation of a PAMdeploy without Training Session Manager aid from an experienced FDO. The philoso- phy of the Payload-assist module The training session manager (TSM) Deploys/Intelligent Computer-Aided consists of two expert systems: an error Training (PD/ICAT) system is to emulate, detection component which compares the the extent possible, the behavior of an assertions made by DeplEx (of both correct experienced FDO devoting his full time and and incorrect actions in a particular con- attention to the training of a novice-- text) with those made by the trainee and proposing challenging training scenar- an error-handling component that decides ios, monitoring and evaluating the actions on the appropriate method of guidance of the trainee, providing meaningful based on the trainee’s skill level. comments in response to trainee errors, responding to trainee requests for infor- Trainee Model mation and hints (if appropriate), and The trainee model is an object-oriented remembering the strengths and weak- data structure which contains a history of nesses displayed by the trainee so that the individual trainee’s interactions appropriate future exercises can be together with summary evaluative data. designed. The model also has a report-generation The PD/ICAT system architecture con- feature that produces a formatted trace of sists of five components and is organized each trainee session and provides the around a common blackboard to facilitate trainee’s supervisor with a high-level communication among the different com- description of the trainee’s current skill ponents (Loftin, Wang, Baffes, and Hua, level and progress. 1988). Training Scenario Generator User Interface The training scenario generator is an The user interface permits the trainee expert system and database that designs to access the same information available to increasingly-complex training exercises him or her in the MCC and serves as a based on the current skill level contained means for the trainee to take actions and in the trainee’s model and on any weak- receive feedback from the training ses- nesses or deficiencies that the trainee has sion manager. The interface mimics, to exhibited in previous interactions. The

106 MORKSHEETHEMU ACTION: Set the sign or the PARAMETER= minus ACTION: Enter the PKM Time PARAMETERS: O, 12, 31, and 40.7

You should have subt, Pact ed, not eddsd, the offset, to injection node t,ime to ¯ st, the pkmLime.

ORBITEREPH TRRGETEPH

Figure 1. A typical screen as seen by a user of the PD/ICAT system. Menus at the upper right allow the trainee to interact with the system. Displays, identical to those used in the MCC, appear in the lower left. A worksheet used by the FDO in completing the task is shown in the lower right. System messages are displayed in the upper left, and error/help messges are provided in a pop-up window near the center of the screen. database serves as a repository for all board also contains a representation of the parameters needed to define a training current trainee action(s) and is a source scenario and includes problems or abnor- of data for the updating of the trainee malities of graded difficulty. The nature model. of this component of the PD/ICAT architecture is more fully developed elsewhere Innovative Features of the (Loftin, Wang and Baffes, 1988). Application System Integration Although a number of intelligent training systems have been developed [for All of the expert system components of example, SOPHIE (Brown, Burton and de PD/ICAT (DeplEx, TSM, and TSG) communi- Kleer, 1982) and STEAMER (Hollan, cate via a common blackboard. The black- Hutchins and Weitzman, 1984)], few, if

107 any, can be said to have been deployed. trainee will catch his or her own The PD/ICAT system has been deployed and mistake. is in use for the training of novice flight ¯ The training scenario generator ex- controllers and for practice and refresh- amines the trainee model and creates ment by experienced personnel. Many a unique scenario for each trainee features of the PD/ICAT system are inno- whenever a new session begins. vative in their own right or are innova- This scenario is built from a database tive in their application to intelligent containing a range of typical pa- training systems: rameters describing the training ¯ PD/ICAT is composed, in part, of four context as well as problems of graded expert systems which cooperate difficulty. Scenarios evolve to through and communicate by means greater difficulty as the trainee of a common blackboard. This demonstrates the acquisition of approach was used to permit the seg- greater skills in solving, the training regation of domain-independent problems. knowledge so that the system archi- ¯ At the conclusion of each session, tecture could be adapted to different the trainee is provided with a training tasks. formatted trace of the session which ¯ Unlike most intelligent tutor- highlights the correct and incorrect ing/training systems, PD/ICAT does actions taken, the time required to not require the trainee to follow a complete the exercise, and the type single correct path to the solution of of assistance provided by the system. a problem. Rather, a trainee is per- In addition, the trainee’s supervisor mitted to select any correct path, as may view a global history of each determined by the scenario context. trainee’s interaction with the system The method used to accomplish this and even generate graphs of trainee flexibility, without generating a performance measured against a combinatorial explosion of solution number of variables. paths, is believed to be unique. ¯ Error detection occurs through the Criteria for a Successfully comparison of the trainee’s actions with those of an expert. In the case Deployed Application of complex actions, the error detec- From the inception of the PD/ICAT tion is made at the highest level to project, a number of factors were recog- avoid confusing the trainee by nized as essential for success at meeting its detecting all errors which propa- objectives and acceptance by its intended gated from the one deemed most sig- audience: nificant. 1. The involvement of the ultimate ¯ Error handling may be accomplished users (or "customers") throughout through the matching of trainee the development process was cer- actions with mal-rules containing tainly the most important factor in errors that are commonly made by PD/ICAT’s successful deployment. novices. In addition, the TSM error- This involvement allowed the handling component decides, based development team to have at hand on the trainee model, what type of the experts in the targeted proce- feedback to give the trainee. Expla- dure and to quickly test competing nations or hints may be detailed for approaches to the solution of spe- novices and quite terse for more cific development problems. experienced personnel. In some 2. The intended audience for the cases, the TSMmay decide not to call training system was asked to pro- attention to the error if there is a vide commentary at each stage of reasonable probability that the the interface development. Thus, at the project’s conclusion, the inter-

108 face had already achieved tacit people but deliver training to only one acceptance by those it was intended person in each position. The ability of a to serve. given trainee to get significant exposure 3. The coding, in a production-rule to a particular process is, therefore, quite system, of those components of limited. The PD/ICAT system, on the other PD/ICAT that would require alter- hand, can provide a trainee with virtually ation as the procedures it was unlimited access to training in a specific designed to teach are altered pro- procedure and insure that the integrated vides built-in maintainability. The simulation environment can be used to AIS has found, after fielding a maximum effect. number of production rule systems, This deployed system has demonstrated that such systems may be easily the capability of intelligent training sys- altered and verified as the task they tems in NASA’s operational environment. were designed to accomplish As a result, a number of similar systems changes. are~under development at JSC and other 4. Sufficient training in production- NASA operational centers (Marshall Space rule coding and detailed documen- Flight Center and Kennedy Space Center). tation were provided to the system’s In addition to the impact of this technol- users so that they were able to pro- ogy on Space Shuttle training, NASA is vide their own long-term support supporting its application to future Space for PD/ICAT. Station training. Since Space Station 5. Although PD/ICAT was developed on training may be a task at least an order of a Symbolics Lisp machine using magnitude larger than current Space ART and Lisp, it has been ported to a Shuttle training, the use of intelligent unix-based workstation using C and training systems may be the only way to CLIPS (a production-rule system, meet Space Station training objectives written in C, and developed by the with the available resources. To this end, AIS at JSC). Such workstations were the Space Station program is supporting already in use by PD/ICAT’s the refinement of the PD/ICAT architec- intended users, and they were not ture into a generic training architecture required to purchase and learn to and the creation of a general-purpose use a different hardware platform. development environment for the rapid 6. An important motivation for man- creation of and adaptation of intelligent agers in the intended user commu- training systems. This latter activity will nity was the ability of PD/ICAT to have a profound impact on the nature of capture the expertise of personnel training, not only within NASA, but who were to be transferred to other within other government agencies, areas. This was a key factor in the industry, and the education establishment. ready availability of the experts and in the management’s support of Deployment Process their dedication of time to this pro- The conception and initial planning ject. for the PD/ICAT project began in July 1986, and initial knowledge acquisition Benefits to NASA was complete by December 1986. The code Training of astronauts and ground- development began in January 1987 and based flight controllers and system engi- the system was demonstrated in March neers is a massive task. The best training 1988. During this latter period three com- and the mechanism for certifying that plete versions of the system were devel- personnel have met training objectives oped before PD/ICAT was accepted. occurs through large-scale, integrated PD/ICAT has been used by both novice simulations. Unfortunately, these simula- FDOs for training and by experienced FDOs tions require the support of hundreds of for practice and refreshment of skills

109 since March 1988. The rehost of PD/ICAT Hartley and Sleeman, 1973. Hartley, J.R. to a unix-based workstation was accom- and Sleeman, D.H., Towards Intelligent plished early in 1989. Teaching Systems, International Journal The development of PD/ICAT required of Man-Machine Studies, 5: 215-236. approximately 3.5 man-years and was Hollan, Hutchins and Weitzman, 1984. accomplished by a mixed team from Hollan, H.D., Hutchins, E.L., and Weitzman, academia (RBL), NASA civil service (LW L. Steamer: An Interactive Inspectable and PB), and private industry (GH). In Simulation-based Training System, AI addition, a number of students contributed Magazine, 5(2): 15-27. to the project during its life. The actual direct cost of the project was approxi- Kearsley, 1987. Artificial Intelligence and mately $120,000 (for the academic and pri- Instruction, ed. Kearsley, G. Reading. MA: vate sector portions). The indirect costs of Addison-Wesley. civil service manpower are more difficult Loftin, Wang and Baffes, 1988. Loftin, R.B., to calculate, but were approximately Wang, L, and Baffes, P., Simulation Sce- $100,000. nario Generation for Intelligent Training Systems. In Proceedings of the Third Arti- Acknowledgments ficial Intelligence and Simulation Work- shop, 69-73. Menlo Park, CA: American The authors wish to acknowledge the Association for Artificial Intelligence. invaluable contributions of expertise from three FDOs: Capt. Wes Jones, USAF; Maj. Loftin, Wang, Baffes and Hua, 1988. Loftin, Doug Rask, USAF(ret.); and Kerry Soileau. R.B., Wang, L., Baffes, P. and Hua, G. An Various students assisted with the knowl- Intelligent Training System for Space edge engineering and coding of portions Shuttle Flight Controllers, Telematics and of the user interface and TSM: Tom Blinn, Informatics, 5: 151-161. Joe Franz, Bebe Ly, Wayne Parrott, and Sleeman and Brown, 1982. Intelligent Chou Pham. Finally, the encouragement Tutoring Systems, eds., Sleeman D. and and guidance of Chirold Epp (Head, ODS) Brown, J.S. London: Academic Press. and Bob Savely (Head, AIS) are gratefully Sleeman, 1982. Sleeman, D.H. Inferring acknowledged. Financial support for this (mal) Rules from Pupils’ Protocols. endeavor has been provided by the Mis- Proceedings of the European Conference sion Planning and Analysis Division, on Artificial Intelligence, 160-164. Orsay, NASA/Johnson Space Center, the NASA France. Office of Space Flight, and, for RBL, by Wenger, 1987. Wenger, E., Artificial NASA/American Society for Engineering Intelligence and Tutoring Systems. Los Education Summer Faculty Fellowships and Altos, CA: Morgan Kaufmann Publishers. a NASA/National Research Council Senior Resident Research Associateship. Yazdani, 1986. Yazdani, M., Intelligent Tutoring Systems Survey, Artificial Intel- References ligence Review, 1: 43-52. Brown, Burton and de Kleer, 1982. Brown, J.S., Burton, R.R., and de Kleer, J., Peda- gogical, Natural Language and Knowledge Engineering Techniques in SOPHIE I, II, and III, in Intelligent Tutoring Systems, 227-282, eds. Sleeman, D. and Brown, J.S. London Academic Press. Carbonell,1970. Carbonell, J.R. AI in CAI: An Artificial Intelligence Approach to CAI, IEEE Transactions on Man-Machine Systems, 11(4): 190-202.

ii0