Strategic behaviour-based reasoning with dynamic, Partial information

Abstract

Behaviour-based agents rely on local perceptions for behaviour selection. However, this can cause problems in dynamic worlds where only partial information is available at any point in time, leading to unstable oscillation between behaviours. While retaining previous perceptual information can alleviate this problem, augmenting the behaviour selection mechanism with a classical approach to belief maintenance is not appropriate in truly dynamic environments, since the environment can change without the agent being aware of it. Using ideas from dynamic belief networks, we augment behaviour selection with information which is temporally degraded: i.e. information is retained from previous percepts but loses certainty, and influence on behaviour, as time passes. We have implemented this approach in the RoboCup simulation domain, using a hierarchical behaviour-based architecture: higher-level behaviours provide the agent with the capability to reason about strategy, but also require non-local information about the environment. We describe a series of controlled experiments which demonstrate that this general-purpose approach to information management addresses the specific problem of unstable behaviour oscillation, while also improving other aspects of the agent’s performance.

The work described was performed while James Westendorp and Paul Scerri were students in the Computer Science Department at RMIT.