Highly Reactive Decision Making: A Game with Time

Abstract

Real-time monitoring calls for decision making capabilities in reaction to observed events. Associative models provide efficiency by matching the observed situation to a recorded pattern equipped with an accurate decision. We rely on a decision tree accounting for the context and temporal chronicles expressing dynamic patterns. In highly reactive domains, i.e. when actions get as frequent as observations, the decision must anticipate the complete recognition of a pattern, comparing possible evolutions. This paper focuses on the on-line decision process, a game against Nature in the general case: a timed game automaton gathers the possible next steps with associated goodness values, and uses an opportunistic algorithm to compute a temporally expressive decision, maximizing its utility, i.e. the chances of winning.