Critic-Driven Voronoi Quantization Distills RL Policies into Explainable Models
A new method, Critic-Driven Voronoi State Partitioning, addresses the performance-interpretability trade-off in distilling deep reinforcement learning policies. Traditional distillation minimizes behavioral distance but ignores action value. This model-agnostic approach partitions a black-box policy into regions optimized by simple models via gradient descent, using the critic value network to iteratively add subpolicies where value is low. The Voronoi quantizer assigns linear functions through nearest neighbor lookups.
Key facts
- Introduces Critic-Driven Voronoi State Partitioning
- Model-agnostic method for distilling deep RL policies
- Partitions black-box policy into regions for simple model optimization
- Uses critic value network to identify regions needing new subpolicies
- Voronoi quantizer assigns linear functions via nearest neighbor lookups
- Addresses performance-interpretability trade-off
- Considers action value, unlike traditional distillation
- Published on arXiv with ID 2605.14897
Entities
Institutions
- arXiv