Predicting Plan Failure by Monitoring Action Sequences and Duration

Abstract

Anticipating failures in agent plan execution is important to enable an agent to develop strategies to avoid or circumvent such failures, allowing the agent to achieve its goal. Plan recognition can be used to infer which plans are being executed from observations of sequences of activities being performed by an agent. In this work, we use this symbolic plan recognition algorithm to find out which plan the agent is performing and develop a failure prediction system, based on plan library information and in a simplified calendar that manages the goals the agent has to achieve. This failure predictor is able to monitor the sequence of agent actions and detects if an action is taking too long or does not match the plan that the agent was expected to perform. We showcase this approach successfully in a health-care prototype system.