Policy Regression for Monitoring Execution in Goal Reasoning Systems

In this paper we investigate the problem of representing and defining policy regression expectations, needed for execution monitoring, when the course of action being monitored is represented as a policy. Until now, the notion of expectation for goal reasoning systems has focused on plans, defined as sequences of actions. This paper presents a formalization to the notion of expectations when the agent is executing policies instead of plans. It also relaxes the notion of policy. Policies are used when acting in probabilistic domains, where executing actions might have multiple outcomes according to some probability distributions.