Learning Causal Models via Progressive Alignment & Qualitative Modeling : A Simulation

Learning causal models is one of the central problems of cognitive science. We describe a simulation of early learning in physical domains from observations that uses progressive alignment and qualitative modeling to derive plausible causal models from observations. We show how protohistories can be created via progressive alignment and used with covariance algorithms to infer causality. The result, a causal corpus described using qualitative representations, can make simple predictions and set the stage for more sophisticated later models. The simulation has been successfully tested with three learning problems, with encouraging results.