Inducing Causal Laws by Regular Inference

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. In this paper, we propose an algorithm to induce the effects of actions from an incomplete domain description and observations after executing action sequences, all of which are represented in the action language $\mathcal{A}$. Our induction algorithm generates effect propositions in $\mathcal{A}$ based on regular inference, i.e., an algorithm to learn finite automata. As opposed to previous work on learning automata from scratch, we are concerned with explanatory induction which accounts for observations from background knowledge together with induced hypotheses. Compared with previous approaches in ILP, an observation input to our induction algorithm is not restricted to a narrative but can be any fact observed after executing a sequence of actions. As a result, induction of causal laws can be formally characterized within action languages.