Causal Generalization in Autonomous Learning Controllers

Abstract

Any machine targeted for human-level intelligence must be able to autonomously use its prior experience in novel situations, unforeseen by its designers. Such knowledge transfer capabilities are usually investigated under an assumption that a learner receives training in a source task and is subsequently tested on another similar target task. However, most current AI approaches rely heavily on human programmers, who choose these tasks based on their intuition. Another largely unaddressed approach is to provide an artificial agent with methods for transferring relevant knowledge autonomously. One step towards effective autonomous generalization capabilities builds on (autonomous) causal modeling and inference processes, using task-independent knowledge representations. We describe a controller that enables an agent to intervene on a dynamical task to discover and learn its causal relations cumulatively from experience. Our controller bootstraps its learning from knowledge of correlation, then removes non-direct-cause correlations – correlations that are due to a common (external) cause, be spurious, or invert cause and effect – through strategic causal interventions, while learning the functions relating a task’s causal variables. The effectiveness of knowledge transfer by the proposed controller is tested through simulation experiments.