Cognitive Architectures and General Intelligent Systems

AI Magazine Volume 27 Number 2 (2006) (© AAAI) Articles Cognitive Architectures and General Intelligent Systems

Pat Langley

Unfortunately, modern artificial intelligence ■ In this article, I claim that research on cognitive ar- has largely abandoned this objective, having chitectures is an important path to the development of general intelligent systems. I contrast this instead divided into many distinct subfields paradigm with other approaches to constructing that care little about generality, intelligence, or such systems, and I review the theoretical commit- even systems. Subfields like computational lin- ments associated with a cognitive architecture. I il- guistics, planning, and computer vision focus lustrate these ideas using a particular architec- their attention on specific components that ture—ICARUS—by examining its claims about underlie intelligent behavior, but seldom show memories, about the representation and organiza- concern about how they might interact with tion of knowledge, and about the performance and each other. Subfields like knowledge represen- learning mechanisms that affect memory struc- tation and machine learning focus on idealized tures. I also consider the high-level programming language that embodies these commitments, draw- tasks like inheritance, classification, and reac- ing examples from the domain of in-city driving. tive control that ignore the richness and com- In closing, I consider ICARUS’s relation to other cog- plexity of human intelligence. nitive architectures and discuss some open issues The fragmentation of artificial intelligence that deserve increased attention. has taken energy away from efforts on general intelligent systems, but it has led to certain types of progress within each of its subfields. Despite this subdivision into distinct commu- nities, the past decade has seen many applica- The Need for General tions of AI technology developed and fielded Intelligent Systems successfully. Yet these systems have a “niche” flavor that differs markedly from those origi- he original goal of artificial intelligence nally envisioned by the field’s early researchers. was the design and construction of com- More broadly based applications, such as hu- putational artifacts that combined many T man-level tutoring systems, flexible and in- cognitive abilities in an integrated system. structable household robots, and believable These entities were intended to have the same characters for interactive entertainment, will intellectual capacity as humans and they were require that we develop truly integrated intelli- supposed to exhibit their intelligence in a gen- gent systems rather than continuing to focus eral way across many different domains. I will on isolated components. refer to this research agenda as aimed at the cre- As Newell (1973) argued, “You can’t play ation of general intelligent systems.

twenty questions with nature and win.” At the However, Newell’s vision for research on in- time, he was critiquing the strategy of experi- tegrated theories of intelligence included more mental cognitive psychologists, who studied than either of these frameworks provides. He isolated components of human cognition with- believed that agent architectures should incor- out considering their interaction. However, porate strong theoretical assumptions about over the past decade, his statement has become the nature of the mind. An architectural design an equally valid criticism of the fragmented na- should change only gradually, as one deter- ture of AI research. Newell proposed that we mines that new structures and processes are re- move beyond separate phenomena and capa- quired to support new functionality. Moreover, bilities to develop complete models of intelli- early design choices should constrain heavily gent behavior. Moreover, he believed that we those made later, producing far more interde- should demonstrate our systems’ intelligence pendence among modules than assumed by ei- on the same range of domains and tasks as han- ther multiagent or blackboard systems. Newell dled by humans, and that we should evaluate (1990) claimed that architectural research is all them in terms of generality and flexibility, about mutual constraints, and its aim should rather than success on a single domain. He also be a unified theory of intelligent behavior, not viewed artificial intelligence and cognitive psy- merely an integrated one. chology as close allies with distinct yet related The notion of a cognitive architecture revolves goals that could benefit greatly from working around this interdependent approach to agent together. This proposal was linked closely to his design. Following Newell’s lead, research on notion of a cognitive architecture, an idea that I such architectures makes commitments about: can best explain by contrasting it with alterna- (1) the short-term and long-term memories tive frameworks. that store the agent’s beliefs, goals, and knowledge; (2) the representation and organization of structures that are embedded in these mem- Three Architectural Paradigms ories; (3) the functional processes that operate Artificial intelligence has explored three main on these structures, including both perfor- avenues to the creation of general intelligent mance and learning mechanisms; and (4) a systems. Perhaps the most widely known is the programming language that lets one construct multi-agent systems framework (Sycara 1998), knowledge-based systems that embody the ar- which has much in common with traditional chitecture’s assumptions. These commitments approaches to software engineering. In this provide much stronger constraints on the con- scheme, one develops distinct modules for dif- struction of intelligent agents than do alterna- ferent facets of an intelligent system, which tive frameworks, and they constitute a compu- then communicate directly with each other. tational theory of intelligence that goes beyond The architecture specifies the inputs/outputs of providing a convenient programming para- each module and the protocols for communi- digm. cating among them, but places no constraints In the next section, I will use one such cog- on how each component operates. Indeed, the nitive architecture—ICARUS—to illustrate each ability to replace one large-scale module with of these commitments in turn. ICARUS is neither another equivalent one is viewed as an advan- the oldest nor the most developed architecture; tage of this approach, since it lets teams devel- some frameworks, like ACT (Anderson 1993) op them separately and eases their integration. and Soar (Laird, Newell, and Rosenbloom One disadvantage of the multi-agent systems 1987), have undergone continual development framework is the need for modules to commu- for more than two decades. However, it will nicate directly with one another. Another par- serve well enough to make the main points, adigm addresses this issue by having modules and its differences from more traditional cogni- read and alter a shared memory of beliefs, tive architectures will clarify the breadth and goals, and other short-term structures. Such a diversity of this approach to understanding the blackboard system (Engelmore and Morgan nature of intelligence. 1989) retains the modularity of the first frame- In discussing ICARUS, I will draw examples work, but replaces direct communication from the domain of in-city driving, for which among modules with an indirect scheme that we have implemented a simulated environ- relies on matching patterns against elements in ment that simplifies many aspects but remains the short-term memory. Thus, a blackboard ar- rich and challenging (Choi et al. 2004). Objects chitecture supports a different form of integra- in this environment include vehicles, for tion than the multiagent scheme, so that the which the positions, orientations, and veloci- former comes somewhat closer to theories of ties change over time, as well as static objects human cognition. like road segments, intersections, lane lines,

34 AI MAGAZINE Articles

Perceptual Buffer Conceptual Belief Memory Memory

Long-Term Short-Term Environment Memories Memories

Skill Goal/Intention Memory Memory

Motor Buffer

Figure 1. Six Long-Term and Short-Term Memories of the ICARUS Architecture.

sidewalks, and buildings. Each vehicle can alter general intelligent systems: (1) cognition is its velocity and change its steering wheel angle grounded in perception and action; (2) con- by setting control variables, which interact cepts and skills are distinct cognitive structures; with realistic laws to determine each vehicle’s (3) long-term memory is organized in a hierar- state. We have implemented ICARUS agents in chical fashion; (4) skill and concept hierarchies other domains, but this is the most complex are acquired in a cumulative manner; and (5) and will serve best to communicate my main long-term and short-term structures have a points. strong correspondence. These ideas further distinguish ICARUS from most other cognitive architectures that have been developed within CARUS The I Architecture the Newell tradition. Again, I will not claim As noted above, ICARUS is a cognitive architec- here that they make the framework superior to ture in Newell’s sense of that phrase. Like its earlier ones, but I believe they do clarify the di- predecessors, it makes strong commitments to mensions that deﬁne the space of candidate ar- memories, representations, and cognitive chitectures. processes. Another common theme is that it incorporates key ideas from theories of human Memories and Representations problem solving, reasoning, and skill acquisi- To reiterate, a cognitive architecture makes a tion. However, ICARUS is distinctive in its con- commitment to the memories that store the cern with physical agents that operate in an ex- content and that control its behavior. These ternal environment, and the framework also must include one or more long-term memories differs from many previous theories by focus- that contain knowledge and procedures, along ing on the organization, use, and acquisition of with one or more short-term memories that hierarchical structures. These concerns have store the agent’s beliefs and goals. The contents led to different assumptions than those found of long-term memories1 change gradually or in early architectures such as ACT and Soar. not at all, whereas short-term elements change Our research on ICARUS has been guided by rapidly in response to environmental condi- five high-level principles about the nature of tions and the agent’s agenda. Some architec-

SUMMER 2006 35 Articles

tures also incorporate sensorimotor memories match, and a :tests field that specifies Boolean that hold information about perceptions and tests it must satisfy. actions; these are updated rapidly as the system Table 1 shows some concepts from the dri- perceives new objects and executes its proce- ving domain. For example, the nonprimitive dures. concept driving-well-in-segment takes three argu- As figure 1 depicts, ICARUS includes six dis- ments: ?self, ?seg, and ?lane. The :percepts field tinct memories, two of each variety. Unlike tra- indicates that the first refers to the agent itself, ditional production-system architectures, the second denotes a road segment, and the which encode all long-term contents as condi- last refers to a lane line associated with the seg- tion-action rules, it has two separate memories, ment. Each structure must match against ele- one for conceptual knowledge and another for ments in the perceptual buffer, which I describe skills or procedures. The framework has two below. The :relations field states that the agent analogous short-term memories, one for the must hold five interconnected beliefs about agent’s beliefs about the environment and an- these entities before it can infer an instance of other for its goals and associated intentions. Fi- the concept. nally, ICARUS has a perceptual buffer that holds Another nonprimitive concept in the table, immediate perceptions of the environment in-rightmost-lane, shows that the :relations field and a motor buffer that contains skills and ac- can also contain negated conditions. These tions it intends for immediate execution. match only when the agent does not hold any ICARUS’s focus on physical settings distin- corresponding belief, with unbound variables guishes it from traditional cognitive architec- like ?anylane being treated as universally quan- tures, including early versions of Soar and ACT, tified. The table also illustrates two primitive although both frameworks have since been ex- concepts, in-lane and in-segment, that refer to tended to interface with external environ- perceived entities and Boolean tests on their at- ments. For example, Laird and Rosenbloom tribute values, but that do not refer to any sim- (1990) report a variant of Soar that controls a pler concepts. ICARUS concepts are very similar physical robot, whereas Byrne (2001) describes to Horn clauses in the programming language ACT-R/PM, which augments ACT-R with per- Prolog (Clocksin and Mellish 1981), although ceptual and motor buffers. However, both the- the syntax differs somewhat. ories focused initially on central cognition and In contrast, long-term skill memory encodes added other modules at a later date, whereas knowledge about ways to act and achieve goals. ICARUS began as an architecture for reactive ex- Each skill clause has a head, which specifies a ecution and places greater emphasis on interac- concept the clause should achieve upon suc- tion with the physical world. cessful completion, and a body with a variety In addition to positing memories, a cogni- of fields. For primitive skill clauses, the body in- tive architecture makes theoretical claims cludes a :start field that describes the situation about the representations used to store infor- in which the agent can initiate the clause, a :re- mation in those memories. Thus, it commits to quires field that must hold throughout the a particular syntax for encoding long-term and skill’s execution (for use in durative skills), and short-term structures. Most frameworks rely on an :actions field that indicates executable ac- formalisms similar to the predicate calculus tions the clause should invoke. For example, that support expression of relational content. Table 2 shows a primitive skill clause for in-in- These build on AI’s central assumption that intersection-for-right-turn, which is considered on- telligent behavior involves the manipulation of ly when one start condition (in-rightmost-lane) symbolic list structures (Newell and Simon holds at the outset and when three require- 1976). An architecture also specifies the man- ments hold throughout execution. ner in which it organizes these structures in Nonprimitive skill clauses differ from primi- memory, along with the way those structures tive ones in that they have no :actions field, are connected across different memories. since they instead have a :subgoals field that For instance, ICARUS represents the contents specifies a set of subgoals the agent should of long-term conceptual memory as Boolean achieve and the order in which this should concepts that encode knowledge about classes happen. Such higher-level clauses also have a of objects and relations among them. Each :start field, but they lack a :requires field, which concept definition includes a head, which spec- is handled by their primitive subskills. Table 2 ifies its name and arguments, and a body, shows two nonprimitive skill clauses for which includes a :percepts field that describes achieving driving-well-in-segment, which have the types and attribute values of observed per- the same percepts but which have slightly dif- ceptual entities, a :relations field that states low- ferent start conditions and subgoals. The table er-level concepts it must match or must not also includes two skill clauses intended to

36 AI MAGAZINE Articles

((driving-well-in-segment ?self ?seg ?lane) :percepts ((self ?self) (segment ?seg) (lane-line ?lane segment ?seg)) :relations ((in-segment ?self ?seg) (in-lane ?self ?lane) (aligned-with-lane-in-segment ?self ?seg ?lane) (centered-in-lane ?self ?seg ?lane) (steering-wheel-straight ?self)))

((in-rightmost-lane ?self ?clane) :percepts ((self ?self) (lane-line ?clane segment ?seg) (segment ?seg)) :relations ((driving-well-in-segment ?self ?seg ?clane) (last-lane ?clane) (not (lane-to-right ?clane ?anylane))))

((in-lane ?self ?lane) :percepts ((self ?self segment ?seg) (lane-line ?lane segment ?seg dist ?dist)) :tests ((> ?dist -10) (<= ?dist 0)))

((in-segment ?self ?seg) :percepts ((self ?self segment ?seg) (segment ?seg)))

Table 1. Some ICARUS Concepts for In-City Driving, with Variables Indicated by Question Marks. achieve the concept in-segment, one a primitive quire this relationship. Moreover, ICARUS skill variant that invokes the action *steer and an- clauses refer to concepts in many of their fields, other that includes a recursive call to itself. Skill thus providing additional organization on the clauses for in-intersection-for-right-turn have a framework’s long-term memories. similar structure, one of which refers to in- ICARUS’s short-term belief memory contains rightmost-lane as a subgoal, which in turn at- instances of defined concepts, which encode tempts to achieve driving-well-in-segment. specific beliefs about the environment that the Clearly, both long-term memories are orga- agent can infer from its perceptions. Each such nized in hierarchical terms, with more complex instance includes the concept name and ob- skills and concepts being defined in terms of jects in the environment that serve as its argu- simpler components. Each hierarchy includes ments. For example, this memory might con- primitive structures at the bottom and specifies tain the instance (in-lane self g601), which increasingly complex structures at higher lev- could follow from the in-lane concept shown els, although direct and indirect recursion can in table 1. Instances of higher-level concepts, also occur. Most cognitive architectures can such as (driving-well-in-segment me g550 g601), model such hierarchical relations, but few raise also take the form of conceptual predicates this notion to a design principle. For example, with objects as their arguments. Table 3 pre- ACT-R lets production rules specify goals in sents some example beliefs from the in-city dri- their left-hand sides and subgoals in their right- ving domain; these clarify that ICARUS’s beliefs hand sides, but the architecture does not re- about its current situation are inherently rela-

SUMMER 2006 37 Articles

((driving-well-in-segment ?self ?seg ?line) :percepts ((segment ?seg) (lane-line ?line) (self ?self)) :start ((steering-wheel-straight ?self)) :subgoals ((in-segment ?self ?seg) (centered-in-lane ?self ?seg ?line) (aligned-with-lane-in-segment ?self ?seg ?line) (steering-wheel-straight ?self)))

((driving-well-in-segment ?self ?seg ?line) :percepts ((segment ?seg) (lane-line ?line) (self ?self)) :start ((in-segment ?self ?seg) (steering-wheel-straight ?self)) :subgoals ((in-lane ?self ?line) (centered-in-lane ?self ?seg ?line) (aligned-with-lane-in-segment ?self ?seg ?line) (steering-wheel-straight ?self)))

((in-segment ?self ?seg) :percepts ((self ?self) (intersection ?int) (segment ?seg)) :start ((last-lane ?line)) :subgoals ((in-intersection-for-right-turn ?self ?int) (in-segment ?self ?seg)))

((in-segment ?self ?endsg) :percepts ((self ?self speed ?speed) (intersection ?int cross ?cross) (segment ?endsg street ?cross angle ?angle)) :start ((in-intersection-for-right-turn ?self ?int)) :actions ((*steer 1)))

((in-intersection-for-right-turn ?self ?int) :percepts ((lane-line ?line) (self ?self) (intersection ?int)) :start ((last-lane ?line)) :subgoals ((in-rightmost-lane ?self ?line) (in-intersection-for-right-turn ?self ?int)))

((in-intersection-for-right-turn ?self ?int) :percepts ((self ?self) (segment ?seg) (intersection ?int) (lane-line ?lane segment ?seg)) :start ((in-rightmost-lane ?self ?lane)) :requires ((in-segment ?self ?seg) (intersection-ahead ?int) (last-lane ?lane)) :actions ((*cruise)))

((in-rightmost-lane ?self ?line) :percepts ((self ?self) (lane-line ?line)) :start ((last-lane ?line)) :subgoals ((driving-well-in-segment ?self ?seg ?line)))

Table 2. Some ICARUS Skills for the In-City Driving Domain.

38 AI MAGAZINE Articles

(current-street me A) (current-segment me g550) (lane-to-right g599 g601) (first-lane g599) (last-lane g599) (last-lane g601) (at-speed-for-u-turn me) (slow-for-right-turn me) (steering-wheel-not-straight me) (centered-in-lane me g550 g599) (in-lane me g599) (in-segment me g550) (on-right-side-of-road-in-segment me) (intersection-behind g550 g522) (building-on-left g288) (building-on-left g425) (building-on-left g427) (building-on-left g429) (building-on-left g431) (building-on-left g433) (building-on-right g287) (building-on-right g279) (increasing-direction me) (buildings-on-right g287 g279)

Table 3. Partial Contents of ICARUS’s Short-Term Conceptual Memory for the In-City Driving Domain. tional in structure, much as the contents of the system pops the stack upon achieving or short-term memories in other architectures like abandoning the current goal. Soar and ACT. Unlike other cognitive architectures, ICARUS However, ICARUS’s perceptual buffer has a also imposes a strong correspondence between somewhat different character. Elements in this the contents of its long-term and short-term memory, which is refreshed on every cycle, de- memories. In particular, it requires that every scribe individual objects that the agent per- short-term element be a specific instance of ceives in the environment. Each element has a some long-term structure. For example, belief type (for example, building or segment), a memory contains instances of defined con- unique name (for example, g425), and a set of cepts that the agent can infer from its environ- attributes with their associated values. Table 4 mental perceptions. Thus, this memory might gives the partial contents of the perceptual contain the instance (in-segment me g550), buffer for one situation that arises in the in-city which it can infer from the in-segment concept driving domain. This includes six lane lines, shown in table 1. The same holds for instances each of which has a length, width, distance, that appear in goal memory, in which an ele- angle, color, and associated segment. The table ment like (driving-well-in-segment me g550 g601) also shows four perceived buildings, each with indicates the agent’s desire to be aligned and an address, street, distance and angle to the centered with respect to a given segment and corner closest to the agent that faces the street, lane line. In fact, ICARUS cannot encode a goal and distance and angle to the other corner. Ad- without a corresponding long-term concept, ditional percepts describe the agent, road seg- and the intentions attached to goals must be ments, sidewalks, an intersection, and a stop- instances of clauses in long-term skill memory. light.2 Note that most attributes take on This suggests a natural approach to modeling numeric values but that some are symbolic. episodic traces that my colleagues and I plan to ICARUS also incorporates a short-term memo- explore in future work. ry for goals and intentions. This contains a pri- ICARUS’s theoretical position contrasts with oritized set of goal stacks, each of which con- those of Soar and ACT-R, which enforce much tains an ordered list of goals, with an entry weaker connections. The latter states that ele- serving as the subgoal for the one below it on ments in short-term memory are active ver- the list. Each goal entry may have an associated sions of structures in long-term declarative skill instance that specifies the agent’s inten- memory, but makes few claims about the relation to execute that skill, once it becomes ap- tion between generalized structures and specif- plicable, in order to achieve the goal. Entries ic instances of them. In both frameworks, pro- may also contain other information about sub- duction rules in long-term memory contain goals that have been achieved previously or generalized patterns that match or alter speciﬁc abandoned. Only the top entry on each goal elements in short-term memory, but ICARUS’s re- stack is accessible to the ICARUS interpreter, but lationship is more constrained. On this dimen- older information can become available when sion, ICARUS comes closer to Schank’s (1982)

SUMMER 2006 39 Articles

(self me speed 5 angle-of-road -0.5 steering-wheel-angle -0.1) (segment g562 street 1 dist -5.0 latdist 15.0) (lane-line g564 length 100.0 width 0.5 dist 35.0 angle 1.1 color white segment g562) (lane-line g565 length 100.0 width 0.5 dist 15.0 angle 1.1 color white segment g562) (lane-line g563 length 100.0 width 0.5 dist 25.0 angle 1.1 color yellow segment g562) (segment g550 street A dist oor latdist nil) (lane-line g6OO length 100.0 width 0.5 dist -15.0 angle -0.5 color white segment g550) (lane-line g6O1 length 100.0 width 0.5 dist 5.0 angle -0.5 color white segment g550) (lane-line g599 length 100.0 width 0.5 dist -5.0 angle -0.5 color yellow segment g550) (intersection g522 street A cross 1 dist -5.0 latdist nil) (building g431 address 99 street A c1dist 38.2 c1angle -1.4 c2dist 57.4 c2angle -1.0) (building g429 address 74 street A c1dist 29.0 c1angle -2.1 c2dist 38.2 c2angle -1.4) (building g425 address 25 street A c1dist 37.8 c1angle -2.8 c2dist 56.9 c2angle -3.1) (building g389 address 49 street 1 c1dist 49.2 c1angle 2.7 c2dist 53.0 c2angle 2.2) (sidewalk g471 dist 15.0 angle -0.5) (sidewalk g474 dist 5.0 angle 1.07) (sidewalk g469 dist -25.0 angle -0.5) (sidewalk g470 dist 45.0 angle 1.07) (stoplight g538 vcolor green hcolor red))

Table 4. Partial Contents of ICARUS’s Perceptual Buffer for the In-City Driving Domain.

theory of dynamic memory, which does not trast, learning processes are responsible for al- meet all of the criteria for a cognitive architec- tering the contents of long-term memory, ei- ture but which he proposed in much the same ther by generating new knowledge structures spirit. In addition, both frameworks share a or by refining and modulating existing struc- commitment to the hierarchical organization tures. In most architectures, the mechanisms of memory and reserve a central role for con- for performance and learning are closely inter- ceptual inference. twined. For example, figure 2 indicates that ICARUS Performance and Learning Processes includes separate performance modules for Besides making theoretical claims about mem- conceptual inference, skill execution, and ories and their contents’ representations, a cog- problem solving, but they operate on many of nitive architecture also commits to a set of the same structures and they build on each processes that alter these contents. These are others’ results in important ways. In particular, described at the level of functional mecha- the problem-solving process is interleaved with nisms, which is more concrete than Newell’s skill retrieval and execution, and both rely (1982) “knowledge level” and more abstract heavily on beliefs produced by the inference than the implementation level of hardware or module to determine their behavior. Further- wetware. Thus, the architecture specifies each more, the hierarchical organization of long- process in terms of an algorithm or procedure term memory plays a central role in each of that is independent of its implementation de- their mechanisms. tails, yet still operates over particular mental Conceptual inference is the architecture’s structures. most basic activity. On each cycle, the system Research on cognitive architectures, like psy- matches concept definitions in long-term chology, generally distinguishes between per- memory against perceptions and beliefs. When formance processes and learning processes. Per- a concept matches, the module adds an in- formance mechanisms utilize structures in stance of that concept to short-term belief long-term memory to interpret and alter the memory, making it available to support other contents of short-term memory, making them inferences. As the left side of ﬁgure 3 depicts, responsible for the generation of beliefs and the system operates in a bottom-up manner, goals. These typically include methods for starting with primitive concepts, which match memory retrieval, pattern matching, skill selec- against percepts, and working up to higher-lev- tion, inference, and problem solving. In con- el concepts, which match against lower-level

40 AI MAGAZINE Articles

Percepual Buffer

Conceptual Conceptual Belief Memory Inference Memory Perception

Skill Retrieval and Selection Environment

Skill Problem Solving Goal/Intention Skill Memory Skill Learning Memory Execution

Motor Buffer

Figure 2. Functional Processes of the ICARUS Architecture and Their Connections to Memories.

concepts. This cascade continues until ICARUS current goal, it carries out actions in the envi- has deduced all beliefs that are implied by its ronment, but when it cannot find such a path, conceptual knowledge base and by its immedi- it invokes a module for means-ends problem ate perceptions. solving (Newell and Simon 1961). This chains In contrast, the skill execution module pro- backward from the goal off either a skill or con- ceeds in a top-down manner, as the right side of cept definition, pushing the result of each rea- figure 3 illustrates. The process starts from the soning step onto a goal stack. The module con- current goal, such as (on-street me A) or (driving- tinues pushing new goals onto the stack until it well-in-segment me ?seg ?lane), and finds applic- finds one it can achieve with an applicable able paths through the hierarchy that terminate skill, in which case it executes the skill and in primitive skill clauses with executable ac- pops the goal from the stack. If the parent goal tions, such as (*steer 1). A skill path is a chain of involved skill chaining, then this leads to exe- skill instances that starts from the agent’s top- level goal and descends the skill hierarchy, uni- cution of its associated skill and achievement fying the arguments of each subskill clause con- of the parent, which is in turn popped. If the sistently with those of its parent. A path is parent goal involved concept chaining, anoth- applicable if the concept instance that corre- er unsatisfied subconcept is pushed onto the sponds to the intention is not satisfied, if the re- goal stack or, if none remain, then the parent is quirements of the terminal (primitive) skill in- popped. This process continues until the sys- stance are satisfied, and if, for each skill instance tem achieves the top-level goal. in the path not executed on the previous cycle, ICARUS’s performance processes have clear the start conditions are satisfied. This last con- similarities to analogous mechanisms in other straint is necessary because skills may take architectures. Conceptual inference plays many cycles to achieve their desired effects, much the same role as the elaboration stage in making it important to distinguish between Soar, which adds inferences to short-term their initiation and their continuation. memory in a deductive, bottom-up manner for When ICARUS’s execution module can find a use in decision making. Selection and execu- path through the skill hierarchy relevant to its tion of skill paths bears a strong resemblance to

SUMMER 2006 41 Articles

Figure 3. ICARUS Conceptual Clauses (left) Are Matched Bottom Up, Starting from Percepts, Whereas Skill Clauses (right) Are Matched Top Down, Starting from the Agent’s Goals. On the left, darker nodes denote matched concept instances; on the right, lighter nodes indicate an applicable path through the skill hierarchy.

the goal-driven, top-down control typically uti- I should reiterate that ICARUS’s various mod- lized in ACT-R systems, although ICARUS uses ules do not proceed in an independent fashion this idea for executing physical actions rather but are constrained by each others’ operation. than cognitive processing, and it traverses Skill execution is influenced by the inference many levels of the skill hierarchy on each deci- process because the former tests against con- sion cycle. The means-ends problem solver op- cept instances produced by the latter. Problem erates much like the one central to Prodigy solving is constrained both by execution, (Minton et al. 1989), except that it interleaves which it uses to achieve subgoals, and by infer- planning with execution, which reflects ence, which lets it determine when they have ICARUS’s commitment to embedding cognition been achieved. Finally, skill learning draws di- in physical agents. rectly on the results of problem solving, which Finally, ICARUS incorporates a learning mod- let it determine the structure of new skills, and ule that creates a new skill whenever problem inferred beliefs, which determine the start con- solving and execution achieve a goal. The new ditions it should place on these skills. Such structure includes the achieved goal as its head, strong interaction is the essence of a cognitive the subgoals that led to the goal as its subskills, architecture that aspires to move beyond inte- and start conditions that differ depending on gration to a uniﬁed theory of intelligence. whether the solution involved chaining off a skill or concept definition. As discussed in Architectures as more detail elsewhere (Langley and Choi, Programming Languages 2006), learning is interleaved with problem Finally, I should note that a cognitive architec- solving and execution, and it occurs in a fully ture typically comes with an associated pro- incremental manner. Skills acquired earlier are gramming language for use in building knowl- available for inclusion in those formed later, edge-based systems. The syntax of this making the learning process cumulative. ICARUS formalism is linked closely to the framework’s shares with Soar and Prodigy the notion of representational assumptions, with knowledge learning from impasses that are overcome in long-term memory corresponding to the through problem solving, but it differs in its program and with initial short-term elements ability to acquire hierarchical skills in a cumu- playing the role of inputs. The language in- lative fashion that builds on earlier structures. cludes an interpreter that can run the program

42 AI MAGAZINE Articles on these inputs, and usually comes archy that is five levels deep. These ter- Newell, and Rosenbloom 1987) are with tracing facilities that let users in- minate in executable actions for many years older and have been used spect the system’s behavior over time. changing speed, altering the wheel an- by far more people. Other well-known In general, the languages associated gle, and depositing packages. We have but more recent cognitive architec- with cognitive architectures are higher used this domain to demonstrate the tures include EPIC (Kieras and Meyer level than traditional formalisms, let- integration of conceptual inference, 1997) and Clarion (Sun, Merrill, and ting them produce equivalent behav- skill execution, problem solving, and Peterson 2001). I will not attempt to ior with much more concise programs. acquisition of hierarchical skills. be exhaustive here, since research in This power comes partly from the There also exist other, more impres- this area has been ongoing since the architecture’s commitment to specific sive, examples of integrated intelligent 1970s, and I can only hope to men- representations, which incorporate systems developed within more estab- tion a representative sample of this ideas from list processing and first-or- lished cognitive architectures. For in- important intellectual movement. der logic, but it also follows from the stance, Tambe et al. (1995) report a sim- However, I should note that the inclusion of processes that interpret ulated fighter pilot, implemented great majority of research efforts on these structures in a specific way. within the Soar framework, that incor- cognitive architectures, including Mechanisms like pattern matching, porates substantial knowledge about those just mentioned, have focused on inference, and problem solving pro- flying missions and that has been used production systems, in which condi- vide many implicit capabilities that repeatedly in large-scale military train- tion-action rules in long-term memo- must be provided explicitly in tradi- ing exercises. Similarly, Trafton et al. ry match against and modify elements tional languages. For these reasons, (2005) describe an ACT-R system, in short-term memory. This paradigm cognitive architectures support far which controls a mobile robot that in- has proven quite flexible and success- more efficient development of soft- teracts with humans in building envi- ful in modeling intelligent behavior, ware for intelligent systems, making ronments having obstacles and occlu- but this does not mean the space of them the practical choice for many ap- sion. These developers have not cognitive architectures lacks other vi- plications. compared directly the lines of code re- able candidates. For reasons given ear- The programming language associ- quired to program such systems within lier, I view ICARUS as occupying a quite ated with ICARUS comes with the syn- a cognitive architecture and within tra- different region of this space, but it tax for hierarchical concepts and ditional programming languages. How- shares features with Minton et al.’s skills, the ability to load and parse ever, I am confident that the higher- Prodigy, which uses means-ends such programs, and commands for level constructs available in ICARUS, analysis to direct learning, and Freed’s specifying the initial contents of Soar, ACT-R, and similar frameworks al- (1998) APEX, which stores complex short-term memories and interfaces low much simpler programs and far skills in a hierarchical manner. Yet the with the environment. The language more rapid construction of intelligent space is large, and we need more sys- also includes an interpreter that han- agents. tematic exploration of alternative dles inference, execution, planning, frameworks that support general intel- and learning over these structures, ligent systems. along with a trace package that dis- Concluding Remarks The field would also benefit from in- plays system behavior on each cycle. I In the preceding pages, I reviewed the creased research on topics that have re- have presented examples of the syntax notion of a cognitive architecture and ceived little attention within tradition- and discussed the operation of the in- argued for its role in developing general cognitive architectures. For instance, terpreter in previous sections, and my al intelligent systems that have the there has been considerable effort on colleagues and I have used this lan- same range of abilities as humans. I al- procedural memory, but much less on guage to develop adaptive intelligent so examined one such architecture— episodic memory, which supports agents in a variety of domains. ICARUS—in some detail. My purpose quite different abilities. Also, most ar- As noted earlier, the most challeng- was to illustrate the theoretical com- chitectural research has focused on ing of these has involved in-city dri- mitments made by cognitive architec- generating the agent’s own behavior, ving. For example, we have construct- tures, including their statements rather than on understanding the ac- ed an ICARUS program for delivering about system memories, the represen- tions of others around it, which is packages within the simulated driving tation of those memories’ contents, equally important. Nor do many cur- environment that includes 15 primi- and the functional processes that op- rent cognitive architectures explain tive concepts and 55 higher-level con- erate on those contents. I also showed the role that emotions might play in cepts, which range from one to six lev- how, taken together, these assump- intelligent systems, despite their clear els deep. These are grounded in tions can support a programming lan- importance to human cognition. perceptual descriptions for buildings, guage that eases the construction of These and many other issues deserve road segments, intersections, lane intelligent agents. fuller attention in future research. lines, packages, other vehicles, and the I should reiterate that ICARUS is nei- Of course, there is no guarantee that agent’s vehicle. The system also incor- ther the most mature or most widely work on unified cognitive architec- porates eight primitive skills and 33 used framework of this sort. Both ACT- tures will lead to computational sys- higher-level skills, organized in a hier- R (Anderson 1993) and Soar (Laird, tems that exhibit human-level intelli-

SUMMER 2006 43 Articles gence. However, recall that, to date, References nich: Oldenbourg KG. Reprinted in Com- we have only one demonstration that puters and Thought, ed. E. A. Feigenbaum Anderson, J. R. 1993. Rules of the Mind. and J. Feldman. New York: McGraw-Hill, such systems are possible—humans Hillsdale, NJ: Lawrence Erlbaum. 1963. themselves—and most research on Byrne, M. D. 2001. ACT-R/PM and Menu Newell, A.; and Simon, H. A. 1976. Com- cognitive architectures, even when it Selection: Applying a Cognitive Architec- puter Science as Empirical Enquiry: Sym- does not attempt to model the details ture to HCI. International Journal of Human- bols and Search. Communications of the of human behavior, is strongly influ- Computer Studies 55(1): 41–84. ACM 19(3): 113–126. enced by psychological findings. At Choi, D.; Kaufman, M.; Langley, P.; Nejati, Nuxoll, A.; and Laird, J. E. 2004. A Cogni- the very least, studies of human cogni- N.; and Shapiro, D. 2004. An Architecture tive Model of Episodic Memory Integrated for Persistent Reactive Behavior. In Pro- tion are an excellent source of ideas for with a General Cognitive Architecture. In ceedings of the Third International Joint Con- how to build intelligent artifacts, and Proceedings of the Sixth International Confer- ference on Autonomous Agents and Multi most cognitive architectures already ence on Cognitive Modeling, 220–225. Mah- Agent Systems, 988–995. New York: ACM incorporate mechanisms with such wah, NJ: Lawrence Erlbaum. Press. origins. Combined with the aim of de- Schank, R. C. 1982. Dynamic Memory. Cam- veloping strong theories of the mind Clocksin, W. F.; and Mellish, C. S. 1981. bridge, U.K.: Cambridge University Press. Programming in Prolog. Berlin: Springer-Ver- and the desire to demonstrate broad Sun, R.; Merrill, E.; and Peterson, T. 2001. lag. generality, this emphasis makes cogni- From Implicit Skills to Explicit Knowledge: Engelmore, R. S.; and Morgan, A. J., eds. tive architectures a viable approach to A Bottom-Up Model of Skill Learning. Cog- 1989. Blackboard Systems. Reading, MA: nitive Science 25(2): 203–244. achieving human-level intelligence. Addison-Wesley. Sycara, K. 1998. Multi-Agent Systems. AI Acknowledgements Freed, M. 1998. Managing Multiple Tasks Magazine 19(2): 79–93. in Complex, Dynamic Environments. In Tambe, M.; Johnson, W. L.; Jones, R. M.; This material is based on research Proceedings of the Fifteenth National Confer- Koss, F.; Laird, J. E.; Rosenbloom, P. S.; and sponsored by DARPA under agreement ence on Artificial Intelligence, 921–927. Men- Schwamb, K. B. 1995. Intelligent Agents numbers HR0011-04-1-0008 and lo Park, CA: AAAI Press. for Interactive Simulation Environments. FA8750-05-2-0283 and by Grant IIS- Kieras, D.; and Meyer, D. E. 1997. An AI Magazine 16(1): 15–39. 0335353 from the National Science Overview of the Epic Architecture for Cog- Trafton, J. G.; Cassimatis, N. L.; Bugajska, Foundation. The U.S. government is nition and Performance with Application M.; Brock, D.; Mintz, F.; and Schultz, A. to Human-Computer Interaction. Human- authorized to reproduce and distribute 2005. Enabling Effective Human-Robot In- Computer Interaction 12(4): 391–438. reprints for governmental purposes teraction Using Perspective-Taking in Ro- notwithstanding any copyright nota- Laird, J. E.; Newell, A.; and Rosenbloom, P. bots. IEEE Transactions on Systems, Man and tion thereon. The views and conclu- S. 1987. Soar: An Architecture for General Cybernetics 25(4): 460–470. sions contained herein are those of the Intelligence. Artificial Intelligence 33(1): 1–64. author and do not necessarily repre- Pat Langley serves as the Director of the In- stitute for the Study of Learning and Exper- sent the official policies or endorse- Laird, J. E.; and Rosenbloom, P. S. 1990. In- tegrating Execution, Planning, and Learn- tise, Consulting Professor of Symbolic Sys- ments, either expressed or implied, of ing in Soar for External Environments. In tems at Stanford University, and Head of DARPA or the U.S. government. Dis- Proceedings of the Eighth National Conference the Computational Learning Laboratory at cussions with John Anderson, Ran- on Artificial Intelligence, 1022–1029. Menlo Stanford’s Center for the Study of Language dolph Jones, John Laird, Allen Newell, Park, CA: AAAI Press. and Information. He has contributed to the David Nicholas, Stellan Ohlsson, and Langley, P.; and Choi, D. 2006. Learning fields of artificial intelligence and cognitive Stephanie Sage contributed to many Recursive Control Programs from Problem science for more than 25 years, having pub- of the ideas presented in this article. Solving. Journal of Machine Learning Re- lished 200 papers and five books on these Dongkyu Choi, Seth Rogers, and search. 7: 493–518. topics including Elements of Machine Learn- ing. Langley is a AAAI Fellow. He was a Daniel Shapiro have played central Minton, S.; Carbonell, J. G.; Knoblock, C. founding editor of the journal Machine roles in the design and implementa- A.; Kuokka, D.; Etzioni, O.; and Gil, Y. Learning. He was program chair for the Sev- tion of ICARUS, with the former devel- 1989. Explanation-Based Learning: A Prob- enteenth International Conference on Ma- oping the driving agent I have used as lem Solving Perspective. Artificial Intelli- chine Learning. His research has dealt with my central example. gence 40(1–3): 63–118. learning in planning, reasoning, language, Newell, A. 1973. You Can’t Play 20 Ques- vision, robotics, and scientific knowledge Notes tions with Nature and Win: Projective discovery, and he has contributed novel Comments on the Papers of This Sympo- 1. An important form of long-term stor- learning methods to the logical, probabilis- sium. In Visual Information Processing ed. age—episodic memory—has received re- tic, and case-based paradigms. His current W. G. Chase. New York: Academic Press. markably little attention within the cogni- research focuses on methods for construct- tive architecture community, although Newell, A. 1982. The Knowledge Level. Ar- ing explanatory process models in scientific Nuxoll and Laird (2004) report some recent tificial Intelligence 18(1): 87–127. domains and on cognitive architectures for efforts in this area. Newell, A. 1990. Unified Theories of Cogni- physical agents. 2. Of course, we might have modeled the tion. Cambridge, MA: Harvard University results of perception at a finer granularity, Press. say at the level of object surfaces or edges, Newell, A.; and Simon, H. A. 1961. GPS, A but the current architecture is agnostic Program That Simulates Human Thought. about such issues. In Lernende Automaten, ed. H. Billing. Mu-

44 AI MAGAZINE