An Evolutionary Strategy for Automatic Hypotheses Generation inspired by Abductive Reasoning

This paper proposes a new evolutionary strategy – called Evolutionary Abduction (EVA) - designed to target a class of problems called Combinatorial Causal Optimization Problems (CCOP). In a CCOP, the goal is to find combinations of causes that best explain or predict an effect of interest. EVA is inspired by abduction, a powerful form of causal inference employed in many artificial intelligence tasks. EVA defines a set of abductive operators to repeatedly construct hypothetical cause-effect instances, and then automatically assesses their plausibility as well as their novelty with respect to already known instances. Experiments confirm that, given a background knowledge, EVA can construct better hypotheses for a given effect, outperforming alternative strategies based on common metaheurstics previously used for CCOP.