Abstract

Recent work on representing action and change has introduced high-level action languages which describe the effects of actions as causal laws in a declarative way. Among such action languages, the language A is the first and the most basic language. In dynamic domains, an agent needs the ability to react against changes of environment and to generate robust plans for a long-term goal, and appropriate representation is necessary for this purpose. In real problems, however, it is difficult to describe complete causal laws for the domain, but it is easier to get observations. In this paper, we propose an algorithm to learn causal laws from an incomplete domain description in the language A, given observations after performing action sequences. Our learning algorithm generates causal laws based on an algorithm to learn finite automata. We also prove the correctness of the learning algorithm. From the result in this work, induction of the effects of actions can now be formally characterized within action languages.