Discovering Conditions for Intermediate Reinforcement with Causal Models

Learning to perform a task in an environment with sparse feedback is a difficult problem. While several approaches for increasing feedback during learning have been taken, these methods suffer from the dependency on human knowledge and engineering to find good solutions. We propose using causal models to increase the amount of feedback that will improve learning. This approach does not require domain-specific human engineering because causal models can be constructed directly from the environment using empirical data. Preliminary experiments and results show causal models can be used to automatically discover conditions for applying intermediate feedback that accelerate learning.