Learning Event Models that Explain Anomalies

摘要

In this paper, we consider the problem of improving the goal-achievement performance of an agent acting in a partially observable, dynamic environment, which may or may not know all events that can happen in that environment. Such an agent cannot reliably predict future events and observations. However, given event models for some of the events that occur, it can improve its predictions of future states by conducting an explanation process that reveals unobserved events and facts that were true at some time in the past. In this paper, we describe the DISCOVERHISTORY algorithm for discovering an explanation for a series of observations in the form of an event history and a set of assumptions about the initial state. When knowledge of one or more event models is not present, we claim that the capability to learn these unknown event models would improve performance of an agent using DISCOVERHISTORY, and provide experimental evidence to support this claim. We provide a description of this problem, and suggest how the DISCOVERHISTORY algorithm can be used in that learning process.