DNS
Determinantal Point Process Based Neural Network Sampler for Ensemble RL
Published at ICML 2022.
Ensemble methods improve stability and performance of RL agents, but training large ensembles is computationally expensive. DNS uses a k-Determinantal Point Process (k-DPP) to sample a maximally diverse subset of neural networks for backpropagation at each training step.
Key result: 50% reduction in computation while outperforming full-ensemble baselines.
Stack: Python, PyTorch, DPP sampling, ensemble RL