Learning-Rate-Free-Reinforcement-Learning

This repository includes the Learning rate-free version of PPO and DQN, using model selection algorithms.

For choosing the model selection strategy, set the 'modsel_alg' parameter to one of the following during the initialization:

• "D3RB"
  • for Doubling Data-Driven Regret Balancing algorithm
• "ED2RB"
  • for Estimating Data-Driven Regret Balancing
• "Classic"
  • for the regret bound balancing algorithm
• "Corral"
• "UCB"
• "Exp3"

Citations

Model Selection implementations are originally from model selection repository.

Reinforcement learning algorithms are modified versions of cleanRL implementations.