🤖 AI Summary
To address the poor interpretability of deep reinforcement learning (DRL) policies, this paper proposes a model-agnostic interpretability distillation framework. First, it adaptively partitions the state space using Voronoi diagrams; then, it distills locally linear policies on each partition. Unlike prior approaches, it imposes no structural assumptions, thus balancing interpretability and representational capacity while preserving policy transparency and achieving performance alignment—or even improvement. The key innovation lies in coupling geometric partitioning with knowledge distillation, endowing local linear models with both theoretical traceability and empirical effectiveness. Experiments on Gridworld and classical control benchmarks demonstrate that the distilled policies exhibit clear decision logic—e.g., piecewise-linear control laws—and achieve average performance gains of 1.2%–3.7% over the original DRL policies, significantly outperforming existing interpretable baselines.
📝 Abstract
Deep Reinforcement Learning is one of the state-of-the-art methods for producing near-optimal system controllers. However, deep RL algorithms train a deep neural network, that lacks transparency, which poses challenges when the controller has to meet regulations, or foster trust. To alleviate this, one could transfer the learned behaviour into a model that is human-readable by design using knowledge distilla- tion. Often this is done with a single model which mimics the original model on average but could struggle in more dynamic situations. A key challenge is that this simpler model should have the right balance be- tween flexibility and complexity or right balance between balance bias and accuracy. We propose a new model-agnostic method to divide the state space into regions where a simplified, human-understandable model can operate in. In this paper, we use Voronoi partitioning to find regions where linear models can achieve similar performance to the original con- troller. We evaluate our approach on a gridworld environment and a classic control task. We observe that our proposed distillation to locally- specialized linear models produces policies that are explainable and show that the distillation matches or even slightly outperforms the black-box policy they are distilled from.