Critic-Driven Voronoi-Quantization for Distilling Deep RL Policies to Explainable Models

Despite many successful attempts at explaining Deep Reinforcement Learning policies using distillation, it remains difficult to balance the performance-interpretability trade-off and select a fitting surrogate model. In addition to this, traditional distillation only minimizes the distance between the behavior of the original and the surrogate policy while other RL-specific components such as action value are disregarded. To solve this, we introduce a new model-agnostic method called Critic-Driven Voronoi State Partitioning, which partitions a black box control policy into regions where a simple class of model can be optimized using gradient descent. By exploiting the critic value network of the original policy, we iteratively introduce new subpolicies in regions with insufficient value, standing in for a measure of policy complexity. The partitioning, a Voronoi quantizer, uses nearest neighbor lookups to assign a linear function to each point in the state space resulting in a cell-like diagram. We validate our approach on several well known benchmarks and proof that this distillation approaches the original policy using a reasonable sized set of linear functions.

Hierarchical Support Vector State Partitioning for Distilling Black Box Reinforcement Learning Policies

We introduce State Vector Space Partitioning (SVSP), a novel method to mimic a black box reinforcement learning policy using a set of human-interpretable subpolicies. By partitioning a distillation dataset of state action pairs with linear support vector machine splits, SVSP constructs a compact and structured representation of the original policy. Our method improves mean return by +7.4\% over previous critic driven state partitioning attempts such as Voronoi State Partitioning (VSP) and +2.8\% over the original TD3 policy, while reducing the number of required subpolicies against VSP by 82.1\%. Our results pave the path towards a more flexible form of distillation where both the decision boundary and surrogate models can be chosen within a margin of the original black box behavior.

Explainable RL Policies by Distilling to Locally-Specialized Linear Policies with Voronoi State Partitioning

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.

Human-Readable Programs as Actors of Reinforcement Learning Agents Using Critic-Moderated Evolution

With Deep Reinforcement Learning (DRL) being increasingly considered for the control of real-world systems, the lack of transparency of the neural network at the core of RL becomes a concern. Programmatic Reinforcement Learning (PRL) is able to to create representations of this black-box in the form of source code, not only increasing the explainability of the controller but also allowing for user adaptations. However, these methods focus on distilling a black-box policy into a program and do so after learning using the Mean Squared Error between produced and wanted behaviour, discarding other elements of the RL algorithm. The distilled policy may therefore perform significantly worse than the black-box learned policy. In this paper, we propose to directly learn a program as the policy of an RL agent. We build on TD3 and use its critics as the basis of the objective function of a genetic algorithm that syntheses the program. Our approach builds the program during training, as opposed to after the fact. This steers the program to actual high rewards, instead of a simple Mean Squared Error. Also, our approach leverages the TD3 critics to achieve high sample-efficiency, as opposed to pure genetic methods that rely on Monte-Carlo evaluations. Our experiments demonstrate the validity, explainability and sample-efficiency of our approach in a simple gridworld environment.