Can Differentiable Decision Trees Learn Interpretable Reward Functions?

There is an increasing interest in learning reward functions that model human intent and human preferences. However, many frameworks use blackbox learning methods that, while expressive, are difficult to interpret. We propose and evaluate a novel approach for learning expressive and interpretable reward functions from preferences using Differentiable Decision Trees (DDTs) for both low- and high-dimensional state inputs. We explore and discuss the viability of learning interpretable reward functions using DDTs by evaluating our algorithm on Cartpole, Visual Gridworld environments, and Atari games. We provide evidence that that the tree structure of our learned reward function is useful in determining the extent to which a reward function is aligned with human preferences. We visualize the learned reward DDTs and find that they are capable of learning interpretable reward functions but that the discrete nature of the trees hurts the performance of reinforcement learning at test time. However, we also show evidence that using soft outputs (averaged over all leaf nodes) results in competitive performance when compared with larger capacity deep neural network reward functions.