Abstraction, Imagery, and Control in Cognitive Architecture.

Cognitive Architecture;Artificial Intelligence;Mental Imagery;Computer Science;Engineering;Science;Computer Science & Engineering

Wintermute, Samuel B.Lewis, Richard L. ;

University of Michigan

关键词: Cognitive Architecture;    Artificial Intelligence;    Mental Imagery;    Computer Science;    Engineering;    Science;    Computer Science & Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/78795/swinterm_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This dissertation presents a theory describing the components of a cognitive architecture supporting intelligent behavior in spatial tasks. In this theory, an abstract symbolic representation serves as the basis for decisions. As a means to support abstract decision-making, imagery processes are also present. Here, a concrete (highly detailed) representation of the state of the problem is maintained in parallel with the abstract representation. Perceptual and action systems are decomposed into parts that operate between the environment and the concrete representation, and parts that operate between the concrete and abstract representations. Control processes can issue actions as a continuous function of information in the concrete representation, and actions can be simulated (imagined) in terms of it. The agent can then derive useful abstract information by applying perceptual processes to the resulting concrete state.This theory addresses two challenges in architecture design that arise due to the diversity and complexity of spatial tasks that an intelligent agent must address. The perceptual abstraction problem results from the difficulty of creating a single perception system able to induce appropriate abstract representations in each of the many tasks an agent might encounter, and the irreducibility problem arises because some tasks are resistant to being abstracted at all.Imagery works to mitigate the perceptual abstraction problem by allowing a given perception system to work in more tasks, as perception can be dynamically combined with imagery. Continuous control, and the simulation thereof via imagery, works to mitigate the irreducibility problem. The use of imagery to address these challenges differs from other approaches in AI, where imagery is considered as an alternative to abstract representation, rather than as a means to it.A detailed implementation of the theory is described, which is an extension of the Soar cognitive architecture. Agents instantiated in this architecture are demonstrated, including agents that use reinforcement learning and imagery to play arcade games, and an agent that performs sampling-based motion planning for a car-like vehicle. The performance of these agents is discussed in the context of the underlying architectural theory. Connections between this work and psychological theories of mental imagery are also discussed.