Cybernetics and Systems, 28:6(1997) 499-520

Autonomous Agents as Embodied AI

by Stan Franklin


This paper is primarily concerned with answering two questions: What are necessary elements of embodied architectures? How are we to proceed in a science of embodied systems? Autonomous agents, more specifically cognitive agents, are offered as the appropriate objects of study for embodied AI. The necessary elements of the architectures of these agents are then those of embodied AI as well. A concrete proposal is presented as to how to proceed with such a study. This proposal includes a synergistic parallel employment of an engineering approach and a scientific approach. It also proposes the exploration of design space and of niche space. A general architecture for a cognitive agent is outlined and discussed.


This essay is motivated by the call for papers for the Cybernetics and Systems' Special issue on Epistemological Aspects of Embodied AI and Artificial Life. Citing "foundational questions concern[ing] the nature of human thinking and intelligence," specific questions are posed, among them the following:

Q1) "Is it necessary for an intelligent system to possess a body...?"
Q2) "What are necessary elements of embodied architectures?"
Q3) "[W]hat drives these systems?"
Q4) "How are we to proceed in a science of embodied systems?"
Q5) "[H]ow is [meaning] related to real objects?"
Q6) "What sort of ontology is necessary for describing and constructing knowledge about systems?"
Q7) "Which ontologies are created within the systems...?

Further, "concrete proposals on how to proceed" with Embodied AI research are encouraged.

The intent here is to speak to each of these questions, with relatively lengthy discussions of Q2 and Q4, and brief responses to the others. And, a concrete proposal will be made on how to proceed. Much of what follows will also apply to Artificial Life research.

Here are my short answers to the above questions, offered as appetizers for the main courses below.

A1) Software systems with no body in the usual physical sense can be intelligent. But, they must be embodied in the situated sense of being autonomous agents structurally coupled with their environment.

A2) An embodied architecture must have at least the primary elements of an autonomous agent, sensors, actions, drives, and an action selection mechanism. Intelligent systems typically must have much more.

A3) These systems are driven by built-in or evolved-in drives and the goals generated from them.

A4) We pursue a science of embodied systems by developing theories of how mechanisms of mind can work, making predictions from the theories, designing autonomous agent architectures that supposedly embody these theories, implementing these agents in hardware or software, experimenting with the agents to check our predictions, modifying our theories and architectures, and looping ad infinitum.

A5) Real objects exist, as objects, only in the "minds" of autonomous agents. Their meanings are grounded in the agent's perceptions, both external and internal.

A6) An ontology for knowledge about autonomous agents will include sensors, actions, drives, action selection mechanisms, and perhaps representations, goals and subgoals, beliefs, desires, intentions, emotions, attitudes, moods, memories, concepts, workspaces, plans, schedules, various mechanisms for generating some of the above, etc. This list does not even begin to be exhaustive.

A7) Each autonomous agent uses it own ontology which is typically partly built-in or evolved-in and partly constructed by the agent.

My concrete proposal on how to proceed includes an expanded form of the cycle outlined in A4 augmented by Sloman's notion of exploration of design space and niche space (1995).

Now for the main courses.

The Action Selection Paradigm of Mind

Classical AI, along with cognitive science and much of embodied AI has developed within the cognitivist paradigm of mind (Varela et al 1991). This paradigm takes as its metaphor mind as a computer program running on some underlying hardware or wetware. It thus sees mind as information processing by symbolic computation, that is rule-based symbol manipulation. Horgan and Tiensen give a careful account of the fundamental assumptions of this paradigm (1996). Serious attacks on the cognitivist paradigm of mind have been mounted from outside by neuroscientists, philosophers and roboticists. (Searle, J. 1980, Edelman 1987, Skarda and Freeman 1987, Reeke and Edelman 1988, Horgan and Tiensen 1989, Brooks 1990, Freeman and Skarda 1990).

Other competing paradigms of mind include the connectionist paradigm (Smolensky 1988, Varela et al 1991, Horgan and Tiensen 1996) and the enactive paradigm (Maturana 1975, Maturana and Varela 1980, Varela et al 1991). The structural coupling invoked in A1 above derives from the enactive paradigm. The connectionist paradigm offers a brain metaphor of mind rather than a computer metaphor.

The action selection paradigm of mind (Franklin 1995), on which this essay is based, sprang from observation and analysis of various embedded AI systems. Its major tenets follow:

AS1) The overriding task of mind is to produce the next action.

AS2) Actions are selected in the service of drives built in by evolution or design.

AS3) Mind operates on sensations to create information for its own use.

AS4) Mind re-creates prior information (memories) to help produce actions.

AS5) Minds tend to be embodied as collections of relatively independent modules, with little communication between them.

AS6) Minds tend to be enabled by a multitude of disparate mechanisms.

AS7) Mind is most usefully thought of as arising from the control structures of autonomous agents. Thus, there are many types of minds with vastly different abilities.

These tenets will guide much of the discussion below as, for example, A5 above will derive from AS3. As applied to human minds, AS4 and AS5 can be more definitely asserted.

An action produces a change of state in an environment (Luck and D'Inverno 1995). But not every such change is produced by an action, for example the motion of a planet. We say that the action of a hammer on a nail changes the environment, but the hammer is an instrument not an actor. The action is produced by the carpenter. Similarly, it's the driver who acts, not the automobile. In a more complex situation it's the user that acts, not the program that produces payroll checks. In an AI setting the user acts with the expert system as instrument. On the other hand, a thermostat acts to maintain a temperature range.

Since actions, in the sense meant here, are produced only by autonomous agents (see below), AS1 leads us to think of minds as emerging from the architectures and mechanisms of autonomous agents. Thus it seems plausible to seek answers to "foundational questions concern[ing] the nature of human thinking and intelligence" by studying the architectures, mechanisms, and behavior of autonomous agents, even artificial agents such as autonomous robots and software agents.

Autonomous Agents

We've spoken several times of autonomous agents. What are they?

An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future (Franklin and Graesser 1997).

And what sorts of entities satisfy this definition? The following figure illustrates the beginnings of a natural kinds taxonomy for autonomous agents (Franklin and Graesser 1997).

With these examples in mind, let's unpack the definition of an autonomous agent.

An environment for a human will include some range of what we call the real world. For most of us, it will not include subatomic particles or stars within a distant galaxy. The environment for a thermostat, a particularly simple robotic agent, can be described by a single state variable, the temperature. Artificial life agents "live" in an artificial environment often depicted on a monitor (e.g. Ackley and Littman 1992). Such environments often include obstacles, food, other agents, predators, etc. Sumpy, a task-specific software agent, "lives" in a UNIX file system (Song, Franklin and Negatu 1996). Julia, an entertainment agent, "lives" in a MUD on the internet (Mauldin 1994). Viruses inhabit DOS, Windows, MacOS, and even Microsoft Word. An autonomous agent must be such with respect to some environment. Such environments can be described in various ways, perhaps even as dynamical systems (Franklin and Graesser 1997). Keep in mind that autonomous agents are, themselves, part of their environments.

Human and animal sensors need no recounting here. Robotic sensors include video cameras, rangefinders, bumpers or antennae with tactile and sometimes chemical receptors (Brooks 1990, Beer 1990). Artificial life agents use artificial sensors, some modeled after real sensors, other not. Sumpy senses by issuing UNIX commands such as pwd or ls. Virtual Mattie, a software clerical agent (Franklin et al, forthcoming) senses only incoming email messages. Julia senses messages posted on the MUD by other users, both human and entertainment agents. Sensor return portions of the environmental state to the agent. Senses can be active or passive. Though all the senses mentioned above were external, internal senses, proprioception, also is part of many agents' design. Some might consider the re-creation of images from memory (see AS4 above) to be internal sensing.

Again, there's no need to discuss actions of human, animal or even robots. Sumpy's actions consist of wandering from directory to directory, compressing some files, backing up others, and putting himself to sleep when usage of the system is heavy. Virtual Mattie, among other things, corresponds with seminar organizers in English via email, sends out seminar announcements and keeps a mailing list updated. Julia wanders about the MUD conversing with occupants. Again, actions can be external or internal, such as producing plans, schedules, or announcements. Every autonomous agent come with a built-in set of primitive actions. Other actions, usually sequences of primitive actions, can also be built in or can be learned.

The definition of an autonomous agent requires that it pursue its own agenda. Where does this agenda come from? Every autonomous agent must be provided with built-in (or evolved-in) sources of motivation for its actions. I refer to these sources as drives (see AS2 above). Sumpy has a drive to compress files when needed. Virtual Mattie has a drive to get seminar announcements out on time. Drives may be explicit or implicit. A thermostat's single drive is to keep the temperature within a range. This drive is hardwired into the mechanism. Sumpy's four drives are as hardwired as that of the thermostat, except that it's done in software. My statement of such a drive describes a straightforward causal mechanism within the agent. Virtual Mattie's six or so drives are explicitly represented as drives within her architecture. They still operate causally, but not in such a straightforward manner. An accounting of human drives would seem a useful endeavor.

Drives give rise to goals that act to satisfy the drives. A goal describes a desired specific state of the environment (Luck and D'Inverno 1995). I picture the motivations of a complex autonomous agent as comprising a forest in the computational since. Each tree in this forest is rooted in a drive which branches to high-level goals. Goals can branch to lower level subgoals, etc. The leaf nodes in this forest comprise the agent's agenda.

Now that we can recognize an autonomous agent's agenda, the question of pursuing that agenda remains. We've arrived at action selection (see AS1 above). Each agent must come equipped with some mechanism for choosing among its possible actions in pursuit of some goal on its agenda. These mechanisms vary greatly. Sumpy is named after it subsumption architecture (Brooks 1990a). One of its layers uses fuzzy logic (Yager and Filev 1994). Some internet information seeking agents use classical AI, say planning (Etzioni and Weld 1994). Virtual Mattie selects her actions via a considerably augmented form of Maes' behavior net (1990). This topic will be discussed in more detail below.

Finally, an autonomous agent must act so as to effect its possible future sensing. This requires that the agent not only be in and a part of an environment, but that it be structurally coupled to that environment (Maturana 1975, Maturana and Varela 1980, Varela et al 1991). (See also A1 above.) Structural coupling, as applied here, means that the agent's architecture and mechanisms must mesh with its environment so that it senses portions relevant to its needs and can act so as to meet those needs.

Having unpacked the definition of autonomous agent, we can think of it as specifying the appropriate objects of study of Embodied AI.


In the early days of AI there was much talk of creating human level intelligence. As the years passed and the difficulties became apparent, such talk all but disappeared as most AI researchers wisely concentrated on producing some small facet of human intelligence. Here I'm proposing a return to earlier goals.

Human cognition typically includes short and long term memory, categorizing and conceptualizing, reasoning, planning, problem solving, learning, creativity, etc. An autonomous agent capable of many or even most of these activities will be referred to as a cognitive agent. (Sloman calls such agents "complete" in one place (19??), and refers to "a human-like intelligent agent (1995) or to "autonomous agents with human-like capabilities" in another (Sloman and Poli 1995). Riegler's dissertation is also concerned with "the emergence of higher cognitive structures"(1994).) Currently, only humans and, perhaps, some higher animals seem to be cognitive agents.

Recently designed mechanisms for cognitive functions as mentioned above include, Kanerva's sparse distributed memory (1988), Drescher's schema mechanism (1988, 1991), Maes' behavior networks (1990), Jackson's pandemonium theory, Hofstadter and Mitchell's copycat architecture (1994, Mitchell 1993), and many others. Many of these do a fair job of implementing some one cognitive function.

The strategy suggested here proposes to fuse sets of these mechanisms to form control structures for cognitive mobile robots, cognitive artificial life creatures, and cognitive software agents. Virtual Mattie is an early example of this strategy. Her architecture extends both Maes' behavior networks and the copycat architecture and fuses them for action selection and perception.

Given a control architecture for an autonomous agent, we may theorize that human and animal cognition works as does this architecture. Since the specification of every control architecture in this way underlies some theory, the strategy set forth in the last paragraph also entails the creation of theories of cognition. For example, the functioning of sparse distributed memory gives rise to a theory of how human memory operates. Such theories, arising from AI, hopefully can help to explain and predict human and animal cognitive activities.

Thus the acronym CAAT, cognitive agents architecture and theory, arises. The CAAT strategy of designing cognitive agent architectures and creating theories from them leads to a loop of tactical activities as follows:

SL1) Design a cognitive agent architecture.
SL2) Implement this cognitive architecture on a computer.
SL3) Experiment with this implemented model to learn about the functioning of the design.
SL4) Use this knowledge to formulate the cognitive theory corresponding to the cognitive architecture.
SL5) From this theory derive testable predictions.
SL6) Design and carry out experiments to test the theory using human or animal subjects.
SL7) Use the knowledge gained from the experiments to modify the architecture so as to improve the theory and its predictions.

We've just seen the science loop of the CAAT strategy, whose aim is understanding and predicting human and animal cognition with the help of cognitive agent architectures. The engineering loop of CAAT aims at producing intelligent autonomous agents (mobile robots, artificial life creatures, software agents) approaching the cognitive abilities of humans and animals. The means for achieving this goal can be embodied in a branch parallel to the sequence of activities described above. The first three items are identical.

EL1) Design a cognitive agent architecture.
EL2) Implement this cognitive architecture on a computer.
EL3) Experiment with this implemented model to learn about the functioning of the design.
EL4) Use this knowledge to design a version of the architecture capable of real world (including artificial life and software environments) problem solving.
EL5) Implement this version in hardware or software.
EL6) Experiment with the resulting agent confronting real world problems.
EL7) Use the knowledge gained from the experiments to modify the architecture so as to improve the performance of the resulting agent.

The science loop and the engineering loop will surely seem familiar to both scientists and engineers. So, why are they included? Because when applied to autonomous agents synergy results from the subject matter. We've seen that each autonomous agent architecture gives rise to (at least one) theory, the theory that says humans and animals do it like this architecture does. Thus the engineering loop can leap out and influence theory. But, theory also constrains architecture. A theory can give rise to (usually many) architectures that implement the theory. Thus gains in the science loop can leap out and influence the engineering loop. Synergy can occur.

This section expands on short answer A4 above. It also constitutes the first part of my proposal on how to proceed with research in embodied AI. The second part will appear below. Now, how do we design cognitive agents?

Design Principles

Not much is known about how to design cognitive agents, though there have been some attempts to build one (Johnson and Scanlon 1987). Theory guiding such design is in an early stage of development. Brustiloni has offered a theory of action selection mechanisms (1991, Franklin 1995), which gives rise to a hierarchy of behavior types. Albus offers a theory of intelligent systems with much overlap to cognitive agents (1991, 1996). Sloman and his cohorts are working diligently on a theory of cognitive agent architectures, with a high level version in place (1995 and references therein). We'll encounter a bit of this theory below. Also, Baars' global workspace model of consciousness (1988), though intended as a model of human cognitive architecture, can be viewed as constraining cognitive agent design. Here we see the synergy between the science loop and the engineering loop in action.

This section proposes design principles, largely unrelated to each other, that have been derived from an analysis of many autonomous agents. They serve to constrain cognitive agent architectures and, so, will contribute to an eventual theory of cognitive agent design.

Drives: Every agent must have built-in drives to provide the fundamental motivation for its actions. This is simply a restatement of A3 and of AS2 above. It's included here because autonomous agent designers tend to hardwire drives into the agent without mentioning them explicitly in the documentation and, apparently, without thinking of them explicitly. An explicit accounting of an agent's drives should help one to understand its architecture and it niche within its environment.

Attention: An agent in a complex environment who has several senses may well need some attention mechanism to help it focus on relevant input. Attending to all input may well be computationally too expensive.

Internal models: Model the environment only when such models are needed. When possible depend on frequent sampling of the environment instead. Modeling the environment is both difficult and computationally expensive. Frequent sampling is typically cheaper and more effective, when it will work at all. This principle has been enunciated several times before (Brooks 1990a, Brustiloni 1991).

Coordination: In multi-agent systems, coordination can often be achieved without the high cost of communication. We often think of coordination of actions as requiring communication between the actors. Many examples show this thought to be a myth. Again, frequent sampling of the environment may serve as well or even better (Franklin forthcoming).

Knowledge: Build as much needed knowledge as possible into the lower level of an autonomous agent's architecture. Every agent requires knowledge of itself and of its environment in order to act so as to satisfy its needs. Some of this knowledge can be learned. Trying to learn all of it can be expected to be computationally intensive even in simple cases (Drescher 1988). The better tack is to hardwire as much needed knowledge as possible into the agent's architecture. This principle has also been enunciated before (Brustiloni 1991).

Curiosity: If an autonomous agent is to learn in an unsupervised way, some sort of more or less random behavior must be built in. Curiosity serves this function in humans, and apparently in mammals. Autonomous agents typically learn mostly by internal reinforcement. (The notion of the environment providing reinforcement is misguided.) Actions in a particular context whose results move the agent closer to satisfying some drive tend to be reinforced, that is made more likely to be chosen again in that context. Actions whose results move the agent further from satisfying a drive tend to be made less likely to be chosen again. In human and many animals, the mechanisms of this reinforcement include pleasure and pain. Every form of reinforcement learning must rely on some mechanism. Random activity is useful when not solution to the current contextual problem is known, and to allow the possibility of improving a known solution. (This principle doesn't apply to observational only forms of learning such as that employed in memory based reasoning (Maes 1994).

Routines: Most cognitive agents will need some means of transforming frequently used sequences of actions into something reactive so that they run faster. Cognitive scientists talk of becoming habituated. Computer scientists like the compiling metaphor. One example is chunking in SOAR (Laird, Newall and Rosenbloom 1987). Another is Jackson's concept demons (Jackson, John V. 1987, Franklin 1995). Agre's dissertation is concerned with human routines (in press).

Brustiloni (1991) gives other such design principles for agents that employ planning, as many cognitive agents must.

High-level Architetures for Cognitive Agents

Several high-level architectures for cognitive agents have been proposed (Albus 1991, Baars 1988, Ferguson 1995, Hayes-Roth 1995, Jackson 1987, Johnson and Scanlon 1987, Sloman 1995). Some of these include descriptions of mechanisms for implementing the architectures, others do not. With the exception of Sloman's, all of these are architectures for specific agents, or for a specific class of agents. Surprisingly, the intersection of all these architectures is rather small. It's like the story of the blind men and the elephant. What you sense depends on your particular viewpoint.

Here we're concerned with answering question Q2 above about the necessary elements of embodied architectures. I'd also like to push further in search of a general architecture for cognitive agents. This architecture should be constrained by the tenets of the action selection paradigm of mind and, as much as possible, by the design principles of the previous section. Ideas for this architecture may be drawn from those referenced in the previous paragraph, and from VMattie's architecture. Hopefully, this architecture will give rise to a theory that serves to kickoff the CAAT strategy outlined above. We'll produce a plan for this architecture by a sequence of refinements beginning with a very simple model.

Computer scientists often partition a computing system into input, processing and output for beginning students.


Figure 1

The corresponding diagram for an autonomous agent might look as follows.


Figure 2

The short answer A2 guides a refinement of Figure 2. While sensors and actions are explicitly present, drives and action selection are not.


Figure 3

Though drives are explicitly represented in Figure 3, they may well appear only implicitly in the causal mechanisms of a particular agent. The diagram in Figure 3 is explicitly guided by AS1 (action selection) and AS2 (drives). AS3 talks about the creation of information (Oyama 1985), which is accomplished partly by perception (Neisser 1993). Perception provides the agent with affordances (Gibson 1979). Glenberg suggests that sets of these affordances allow the formation of concepts and the laying down of memory traces (to appear). Note that we've split perception off from action selection. In further refinements of the architecture, action selection must be interpreted less and less broadly.

Figure 4

AS4 leads to another refinement with the addition of memory. We will include both long-term memory and short-term memory (workspace). Though only one memory and one workspace will be shown in Figure 5, multiple specialized memories and workspaces may be expected in the architectures of complex autonomous agents. VMattie's architecture contains two of each, one set serving perception.

Figure 5

At this point the action selection paradigm of mind give us only general guidance: employ multiple, independent modules (AS5) and allow for disparate mechanisms (AS5). Thus we turn to design principles. The Attention Principle points to an attention mechanism or relevance filter. Note that attention will depend on context and, ultimately, on the strength and urgency of drives.


Figure 6

Though the Internal Models principle warns against over doing it, some internal modeling of the environment will be needed to allow for expectations (important to perception, for instance), planning, problem solving, etc. This principle also points to the distinction between reactive and deliberative action selection (see for example Sloman 1995). Deliberative actions are selected with the help of internal models, planners, schedulers, etc. These models use internal representations in the strict sense of the word, that is, they are consulted for their content. Internal states that play a purely causal role without such consultations are not representations in this sense (Franklin 1995 Chapter 14). Reactive actions are exemplified by reflexes and routines (Agre, in press). They are arrived at without such consultation. Brustiloni's instinctive and habitual behaviors would be reactive while his problem solving and higher behaviors would be deliberative (1991). Deliberative mechanisms such as planners and problem solvers may well require their own memories and workspaces not shown in the figure.

Figure 7

The Coordination Principle warns us against unnecessary communication. Still, for a cognitive agent in a society of other such, communication may well be needed. It is sufficiently important that some people include it in their definition of an agent (Wooldridge and Jennings 1995). VMattie communicates with humans by email. Her understanding of incoming messages is part of perception. Her composition of outgoing messages are brought about by deliberative behaviors. Independent modules for understanding messages and for composing messages must be part of a general cognitive agent architecture.

Figure 8

The Knowledge Principle brings up two issues: building knowledge into the reactive behaviors, and learning. By definition, knowledge is built into reactive behaviors casually through their mechanisms, rather than declaratively by means of representations. This doesn't show up in the diagrams.

The other issue brought up by the Knowledge Principle is learning, which is critical to many autonomous agents coping with complex, dynamic environments, and must be included in a general cognitive agent architecture. Learning, itself, is quite complex. Thomas (1993) lists eight different types of learning as follows: 1) habituation-sensitization, 2) signal learning, 3) stimulus-response learning, 4) chaining, 5) concurrent discriminations, 6) class concepts: absolute and relative, 7) relational concepts I: conjunctive, disjunctive, conditional concepts, 8) relational concepts II: biconditional concepts. He uses this classification as a scale to measure the abilities of animals to learn. Perhaps the same or a similar scale could be used for autonomous agents.

Thomas' classification categorizes learning according to the sophistication of the behavior to be learned. One might also classify learning according to the method used. Maes lists four such methods: memory based reasoning, reinforcement learning, supervised learning, and learning by advice from other agents (1994). Drescher's concept learning (1988) and Kohonen's self-organization (1984) are other methods. The Curiosity Principle is directly concerned with reinforcement learning, while the Routines Principle speaks of compiling or chunking sequences of actions, again a form of learning.

Figure 9

The limitations of working with an almost planer graph become apparent in Figure 9. Learning mechanisms should also connect to drives and to memory, essentially to everything. And, there are other connections that need to be included, or to run in both directions.

Well, we've run out of our sources of guidance, both action selection paradigm tenets and design principles. Are we then finished? By no means. Our general architecture for a cognitive agent is still in its infancy. Much is missing.

Our agent's motivations are restricted to drives and the goal trees that are grown from them. We haven't even mentioned the goal-generators that grow them. We also haven't discussed other motivational elements, such as moods, attitudes, and emotions which can influence action selection. For such a discussion, see the work of Sloman (1979, 1994, Sloman and Croucher 1981). Perception can have a quite complex architecture of its own (Marr 1982, Sloman 1989, Kosslyn and Koeing 1992), including workspaces and memories. Each sensory modality will require its own unique mechanisms, as will the need to fuse information from several modalities. Each of the deliberative mechanisms will have its own architecture, often including workspaces and memories. Each will have its own connections to other modules. For instance, some set of them may connect in parallel to perception and action selection (Ferguson 1995). Similarly, each of the various learning mechanisms will have its own architecture, connecting in unique ways to other modules. The relationship between sensing and acting, where acting facilitates sensing, isn't yet specified in the architecture. And, the internal architecture of the action selection module itself hasn't be discussed. Finally, there's a whole other layer of the architecture missing, what Minsky calls the B-brain (Minsky 1985), and Sloman calls the meta-management layer (Sloman 1995). This layer watches what's going on in other parts of our cognitive agent's mind, keeps it from oscillating, improves it strategies, etc. And after all this, are we through? No. There's Baars' notion of a global workspace that broadcasts information widely in the system, and allows for the possibility of consciousness (1988). There seems to be no end.

As you can see, the architecture of cognitive agents is the subject, not for an article, but for a monograph, or a ten-volume set. Question Q2 about the necessary elements of embodied architectures is not an easy one if, as I do, you take cognitive agents to be the proper objects of study for embodied AI.

A Paradigm for Embodied AI

So, how should embodied AI research proceed? Here's a "concrete proposal."

Study cognitive agents. The contention here is that a holistic view is necessary. Intelligence cannot be understood piecemeal. That's not to say that projects picking off a piece of intelligence and studying its mechanisms aren't valuable. They often are. The claim is that they are not sufficient, even in the aggregate.

Follow the CAAT strategy. Running the engineering loop and the science loop in parallel will enable the synergy between. This will mean making common cause with cognitive scientists and cognitive neuroscientists.

Explore design space and niche space (Sloman 1995). Strive to understand not only individual agent architectures, but the space of all such architectures. This means exploring, classifying, and theorizing at a higher level of abstraction. Each agent occupies a particular niche in its environment. Explore, in the same way, the space of such niches and the architectures that are suitable to them.


Ackley, David, and Littman, Michael, (1992). "Interactions Between Learning and Evolution." In Christopher Langton et al., ed., Artificial Life II. Redwood City, Calif.: Addison-Wesley 407-509.

Agre, Philip E. (in press). The Dynamic Structure of Everyday Life. Cambridge: Cambridge University Press.

Albus, J.S. (1991). "Outline for a Theory of Intelligence," IEEE Transactions on Systems, Man and Cybernetics, Vol. 21, No. 3, May/June.

Albus, J.S. (1996). "The Engineering of Mind," Proceedings of the Fourth International Conference on Simulation of Adaptive Behavior: From Animals to Animats 4, Cape Code, MA, September.

Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press.

Beer, R. D. (1990). Intelligence as Adaptive Behavior. Boston: Houghton Mifflin.

Brooks, R. A. (1990). "Elephants Don't Play Chess." in Pattie Maes, ed., Designing Autonomous Agents. Cambridge, MA: MIT Press.

Brooks, R. A. (1990a). "A Robust Layered Control System for a Mobile Robot." in P. H. Winston, ed., Artificial Intelligence at MIT, vol. 2. Cambridge, MA: MIT Press.

Brustoloni, Jose C. (1991). "Autonomous Agents: Characterization and Requirements." Carnegie Mellon Technical Report CUM-CS-91-204. Pittsburgh: Carnegie Mellon University.

Drescher, G. L. (1988). "Learning from Experience Without Prior Knowledge in a Complicated World." Proceedings of the AAAI Symposium on Parallel Models. AAAI Press.

Kosslyn and Koeing 1992) Made-up Minds. MIT Press.

Edelman, G. M. (1987). Neural Darwinism: The Theory of Neuronal Group Selection. New York: Basic Books.

Etzioni, Oren and Weld, Daniel (1994). "A Softbot-Based Interface to the Internet." Communications of the ACM, (37) 7: 72-79.

Ferguson, I. A. (1995). "On the role of DBI modeling for integrated control and coordinated behavior in autonomous agents." Applied Artificial Intelligence, 9(4).

Foner, L. N., and Maes, Pattie. (1994). "Paying Attention to What's Important: Using Focus of Attention to Improve Unsupervised Learning." Proceedings of the Third International Conference on the Simulation of Adaptive Behavior, Brighton, England.

Franklin, Stan (1995). Artificial Minds. Cambridge, MA: MIT Press.

Franklin, Stan (forthcoming). "Coordination Without Communication."

Franklin, Stan and Graesser, Art (1997) "Is it an Agent, or just a Program?: A Taxonomy for Autonomous Agents," Proceedings of the Agent Theories, Architectures, and Languages Workshop, Berlin: Springer Verlag, 193-206.

Franklin, Stan, Graesser, Art, Olde, B., Song, H., and Negatu, A. (forthcoming)
"Virtual Mattie-an Intelligent Clerical Agent."

Freeman, W. J., and Skarda, C. (1990). "Representations: Who Needs Them?" In J. L. McGaugh, et al., eds., Brain Organization and Memory Cells, Systems, and Circuits. New York: Oxford University Press.

Gibson, J. J., (1979). The Ecological Approach to Visual Perception. Boston: Houghton Mifflin.

Glenberg, A. M. (to appear). "What Memory Is For." Behavioral and Brain Sciences.

Hayes-Roth, B. (1995). "An architecture for adaptive intelligent systems." Artificial Intelligence, 72 329-365.

Hofstadter, D. R. and Mitchell, M. (1994). "The Copycat Project: A model of mental fluidity and analogy-making." In Holyoak, K.J. & Barnden, J.A. (Eds.) Advances in connectionist and neural computation theory. Vol. 2: Analogical Connections. Norwood, N.J.: Ablex.

Horgan, T. and Tiensen, J. (1989). "Representation without Rules." Philosophical Topics, 17 (Spring): 147-74.

Horgan, T. and Tiensen, J. (1996). Connectionism and the Philosophy of Psychology. Cambridge, MA: MIT Press.

Jackson, John V. (1987). "Idea For A Mind." SIGART Newsletter, July
No. 101 23-26

Johnson, M. and Scanlon, R. (1987) "Experience with a Feeling-Thinking Machine", Proceedings of the IEEE First International Conference on Neural Networks, San Diego. 71-77.

Kanerva, P. (1988). Sparse Distributed Memory. MIT Press.

Kohonen, T. (1984). Self-Organization and Associative Memory. Berlin: Springer Verlag.

Kosslyn, S. M. and Koeing, O. (1992). Wet Minds. New York: The Free Press.

Laird, John E., Newall, Allen, and Rosenbloom, Paul S. (1987). "SOAR: An Architecture for General Intelligence." Artificial Intelligence, 33: 1-64.

Luck, M. and D'Inverno, M. (1995). "A Formal Framework for Agency and Autonomy." Proceedings of the International Conference on Multiagent Systems. 254-260.

Maes, Pattie (1990). "How to do the right thing." Connection Science. (1):3.

Maes, Pattie (1994). "Agents that Reduce Work and Information Overload." Communications of the ACM, Vol. 37, No. 7. 31-40, 146, ACM Press, July.

Marr, David (1982). Vision. San Francisco: W. H. Freeman.

Maturana, H. R. (1975). "The Organization of the Living: A Theory of the living Organization." International Journal of Man-Machine Studies, 7:313-32.

Maturana, H. R. and Varela, F. (1980). Autopoiesis and Cognition: The Realization of the Living. Dordrecht, Netherlands: Reidel.

Mauldin, M. L. (1994). "Chatterbots, Tinymuds, And The Turing Test: Entering The Loebner Prize Competition." Proceedings of AAAI.

Minsky, Marvin (1985). Society of Mind. New York: Simon and Schuster.

Mitchell, Melanie (1993). Analogy-Making as Perception, Cambridge MA: The MIT Press.

Neisser, Ulric (1993). "Without Perception, There Is No Knowledge: Implications for Artificial Intelligence." in R. G. Burton, ed., Natural and Artificial Minds, State University of New York Press.

Oyama, Susan (1985). The Ontology of Information. Cambridge: Cambridge University Press.

Reeke, G. N., Jr., and Edelman, G. M. (1988). "Real Brains and Artificial Intelligence." Daedalus, Winter, 143-73. Reprinted in S. R. Graubard, ed., The Artificial Intelligence Debate, Cambridge, MA: MIT Press.

Riegler, Alexander (1994), "Constructivist Artificial Life," In: Hopf, J. (ed.) Proceedings of the 18th German Conference on Artificial Intelligence (KI-94) Workshop "Genetic Algorithms within the Framework of Evolutionary Computation" Max-Planck-Institute Report No. MPI-I-94-241.

Searle, J. (1980). "Minds, Brains, and Programs." Behavioral and Brain Sciences 3:417-58.

Skarda, C. and Freeman, W. J. (1987). "How Brains Make Chaos in Order to Make Sense of the World." Behavioral and Brain Sciences 10, 2:161-95.

Sloman, Aaron (1979). "Motivational and Emotional Controls of Cognition." reprinted in Model of Thoughts, Yale University Press. 29-38.

Sloman, Aaron (1989). "On Designing a Visual System: Towards a Gibsonian computational Model of Vision." Journal of Experimental and Theoretical AI. 1,7 289-337.

Sloman, Aaron (1994). "Computational modeling of motive-management processes." Proceedings of the Conference of the International Society for Research in Emotions. Cambridge, N. Frijda, ed. ISRE Publications, 344-8.

Sloman, Aaron (1995). "Exploring Design Space and Niche Space." Proceedings 5th Scandinavian Conf on AI, Trondheim May 1995, Amsterdam: IOS Press.

Sloman, Aaron and Croucher, M. (1981). "Why Robots will have Emotions." Proc. 7th Int. Joint Conf. on AI. Vancouver.

Sloman, Aaron and Poli, Riccardo (1995). "SIM_AGENT: A Toolkit for Exploring Agent Designs." in M. Woolridge, et al eds., Intelligent Agents, Vol. II (ATAL-95). Springer-Verlag 392-407.

Song, H., Franklin, Stan and Negatu, A. (1996), "SUMPY: A Fuzzy Software Agent." in ed. F. C. Harris, Jr., Intelligent Systems: Proceedings of the ISCA 5th International Conference (Reno Nevada, June 1996) International Society for Computers and Their Applications - ISCA, 124-129.

Thomas, Roger K. (1993). "Squirrel Monkeys, Concepts and Logic." in R. G. Burton, ed., Natural and Artificial Minds, State University of New York Press.

Varela, F.J., Thompson, E. and Rosch, E. (1991). The Embodied Mind. Cambridge, MA: MIT Press.

Wooldridge, Michael and Nicholas R. Jennings (1995), "Agent Theories, Architectures, and Languages: a Survey," in Wooldridge
and Jennings Eds., Intelligent Agents, Berlin: Springer-Verlag, 1-22

Yager, R. R. and Filev, D. P. (1994). Essentials of Fuzzy Modeling and Control. New York: John Wiley & Sons.