Finding causal theories quickly enough - building a responsive Apperception Engine

I describe the implementation of a responsive Apperception Engine, based on the original design described in “Making sense of sensory input” by Richard Evans et al, for use in a distributed cognitive architecture targeting simple robots. An Apperception Engine applies symbolic machine learning to produce causal theories from short sequences of observations. The causal theories are small, human-readable logic programs that model an environment’s latent generative processes and predict incoming observations. Finding these causal theories quickly is a difficult problem due to the size of the search space. I built an Apperception Engine that can infer good enough causal theories quickly enough to be useful to a robot actively engaging with its environment.