Defense Event

Abstraction, Imagery, and Control in Cognitive Architecture

Samuel Wintermute

Friday, July 23, 2010
1:30pm - 3:30pm
3725 Beyster Bldg.

Add to Google Calendar

About the Event

This dissertation considers the question of what components should be included in a cognitive architecture in order to support intelligent behavior in spatial tasks. An agent can benefit substantially from using abstract internal representations of the structure of these tasks, where unnecessary detail is eliminated. However, due to the diversity and complexity of spatial tasks that an intelligent agent must address, there are significant challenges inherent in designing a general-purpose architecture able to use abstract representations. Here, two of these challenges are examined in detail. The perceptual abstraction problem arises due to 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. To address these challenges, a theory for the basic components of a cognitive architecture is described. In addition to an abstract representation used as a basis for decision-making, this theory includes a concrete (highly detailed) representation of the state of world. Controllers that generate outputs as a continuous function of the content of the concrete representation are present, partially mitigating the irreducibility problem. Imagery capability is also present, where internal simulations in terms of the concrete representation are used to derive abstract information about the consequences of potential actions. Imagery works to mitigate the perceptual abstraction problem, and, via the simulation of continuous control, also works to mitigate the irreducibility problem. This theory leads to a number of benefits that allow for improved performance in particular tasks, and that allow the architecture to address a wider range of tasks. A detailed implementation of the theory is described, which is an extended version 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 benefits of the theory. Connections between this work and psychological theories of mental imagery are also discussed.

Additional Information

Sponsor(s): J. Laird

Open to: Limited