In this paper, we examine the motivations for research on cognitive architectures and review some candidates that have been explored in the literature. After this, we consider the capabilities that a cognitive architecture should support, some properties that it should exhibit related to representation, organization, performance, and learning, and some criteria for evaluating such architectures at the systems level. In closing, we discuss some open issues that should drive future research in this important area. Key words: cognitive architectures, intelligent systems, cognitive processes 1

Abstract. Endowing an intelligent agent with an episodic memory affords it a multitude of cognitive capabilities. However, providing efficient storage and retrieval in a task-independent episodic memory presents considerable theoretical and practical challenges. We characterize the computational issues bounding an episodic memory. We explore whether even with intractable asymptotic growth, it is possible to develop efficient algorithms and data structures for episodic memory systems that are practical for real-world tasks. We present and evaluate formal and empirical results using Soar-EpMem: a task-independent integration of episodic memory with Soar 9, providing a baseline for graph-based, taskindependent episodic memory systems. 1

Symbolic AI systems typically have difficulty reasoning about motion in continuous environments, such as determining whether a cornering car will clear a close obstacle. Bimodal systems, integrating a qualitative symbolic system with a quantitative diagram-like spatial representation, are capable of solving this sort of problem, but questions remain of how and where knowledge about fine-grained motion processes is represented, and how it is applied to the problem. In this paper, we argue that forward simulation of motion is an appropriate method, and introduce continuous motion models to enable this simulation. These motion-specific models control behavior of objects at the spatial level, while general mechanisms in the higher qualitative level control and monitor them. This interaction of low- and high-level activity allows for problem solving that is both precise in individual problems and general across multiple problems. In addition, this approach allows perception and action mechanisms to be reused in reasoning about hypothetical motion problems and abstract non-motion problems, and points to how symbolic AI can become more grounded in the real world. We demonstrate implemented systems that solve problems in diverse domains, and connections to action control are discussed.

This paper brings together work in modeling episodic memory and reinforcement learning. We demonstrate that is possible to learn to use episodic memory retrievals while simultaneously learning to act in an external environment. In a series of three experiments we investigate learning what to retrieve from episodic memory and when to retrieve it, learning how to use temporal episodic memory retrievals, and learning how to build cues that are the conjunctions of multiple features. Our empirical results demonstrate that it is computationally feasible to learn to use episodic memory in all three experiments, and furthermore, that learning to use internal episodic memory accomplishes tasks that reinforcement learning alone does not. These experiments also expose some important interactions that arise between reinforcement learning and episodic memory.

This paper represents a first step in attempting to engage the research community in discussions about evaluation of human-level intelligent systems. First, we discuss the challenges of evaluating human-level intelligent systems. Second, we explore the different types of claims that are made about HLI systems, which are the basis for confirmatory evaluations. Finally, we briefly discuss a range of experimental designs that support the evaluation of claims.

As cognitive architectures become ever more ambitious in the range of phenomena they are to assist in producing and modeling, there is increasing pressure for diversity in the mechanisms they embody. Yet uniformity remains critical for both elegance and extensibility. Here, the search for uniformity is continued, but shifted downwards in the cognitive hierarchy to the implementation level. Factor graphs are explored as a promising core, with initial steps towards a reimplementation of Soar. The ultimate aim is a uniform implementation level for cognitive architectures affording both heightened elegance and expanded coverage.

Predicting the results of one’s own actions is a powerful cognitive capability that can aid in determining which action to take in a given situation. In this paper, we describe a taskindependent framework based on the Soar cognitive architecture in which rules, episodic memory, semantic memory, mental imagery, and task decomposition are available for predicting an action’s consequences. We include results from two domains and make predictions for human behavior based on these results.

Abstract not found

This paper presents an approach for integrating action in the world with general symbolic reasoning. Instead of working with task-specific symbolic abstractions of continuous space, our system mediates action through a simple spatial representation. Low-level action controllers work in the context of this representation, and a high-level symbolic system has access to it. By allowing actions to be spatially simulated, general reasoning about action is possible. Only very simple task-independent symbolic abstractions of space are necessary, and controllers can be used without the need for symbolic characterization of their behavior. We draw parallels between this system and a modern robotic motion planning algorithm, RRT. This algorithm is instantiated in our system, and serves as a case study showing how the architecture can effectively address real robotics problems.

In many spatial problems, it can be difficult to create a state representation that is abstract enough so that irrelevant details are ignored, but also accurate enough so that important states of the problem can be differentiated. This is especially difficult for agents that address a variety of problems. A potential way to resolve this difficulty is by using two representations of the spatial state of the problem: one abstract and one concrete, along with internal (imagery) operations that modify the concrete representation based on the contents of the abstract representation. In this paper, we argue that such a system can allow plans and policies to be expressed that can better solve a wider class of problems than would otherwise be possible. An example of such a plan is described. The theoretical aspects of what imagery is, how it differs from other techniques, and why it provides a benefit are explored.