Maximising expected utility for behaviour arbitration

Abstract

Utility fusion is presented as an alternative means of action selection which ameliorates both the bottlenecks of centralised systems and the incoherence of distributed systems. In this approach, distributed behaviours indicate the utility of possible world states, along with their associated uncertainty. A centralised arbiter then combines these utilities and probabilities to determine a Pareto-optimal action based on the maximisation of expected utility. Utility theory provides a Bayesian framework for explicitly representing and reasoning about uncertainty within the action selection process. In addition, the construction of a utility map allows the arbiter to model and compensate for the dynamics of the system; experimental results verify that the resuhing system provides significantly greater stability.