Last updated 1 November, 2006; originally from 1996

Action Selection, Control State and Reactive, Behavior-Based AI

One contribution of my approach to behavior-based AI has been to separate the organization of control from the flow of data --- separating when a behavior should execute from how . When is determined by action selection, how by ordinary code organized in behavior libraries.

Within the flow of control, I have added mechanisms for focusing attention and maintaining decision state through time. This reduces the combinatorics of the problem of action selection and increases behavior coherence. On the other hand, it reduces reactiveness, since not all behaviors are actively expressing actions at any one time, and perception is a part of behavior. Hierarchy and Sequence vs. Full Parallelism in Reactive Action Selection Architectures, (from The Sixth International Conference on the Simulation of Adaptive Behavior (SAB2000)) demonstrates that this does not necessarily lead to a reduction of performance, even in highly dynamic environments.

I'm hardly the only person working in this area, see this page of related research for some of the others. What makes my approach special is:

For a synopsis of my action selection mechanism, see my POSH web page, for my entire design approach see the Behaviour Oriented Design (BOD) page. Significantly more details are available in my publications.

Earlier thoughts on these problems: If you think this work is heresy, and that everything can be done in a purely reactive system, you might want to read The Use of State in Intelligent Control . This compares Shakey and Genghis, and demonstrates the necessity of using control state even in the simplest reactive system. I also have a web page discussing the utility of goals, which documents the motivation for BOD and POSH.

page author: Joanna Bryson