Reinforcement Learning

SAC and Extensions ★ 295 SAC with PER, Emphasizing Recent Experience, Munchausen RL, D2RL, and parallel environments. CQL ★ 147 Conservative Q-Learning for offline RL. Includes DQN-CQL and SAC-CQL variants for discrete and continuous action spaces. DQN Atari Agents ★ 122 Modular DQN agents for Atari: DDQN, Dueling DQN, Noisy DQN, C51, Rainbow, and DRQN. IQN and Extensions ★ 93 Implicit Quantile Networks for distributional RL with PER, noisy layers, N-step bootstrapping, and dueling architecture. Deep RL Algorithm Collection ★ 80 Comprehensive collection of deep RL algorithm implementations in PyTorch. Upside-Down RL ★ 78 Implementation of Schmidhuber's Upside-Down RL — supervised learning on desired rewards and time horizons. SAC Discrete ★ 55 Soft Actor-Critic adapted for discrete action spaces. Implicit Q-Learning ★ 44 IQL for offline RL — learns optimal policies without querying out-of-distribution actions. Munchausen RL ★ 44 M-DQN and M-IQN: adding scaled log-policy to the reward for implicit KL regularization. FQF and Extensions ★ 34 Fully Parameterized Quantile Function for distributional RL with N-step, PER, noisy layers, and dueling. QR-DQN ★ 29 Distributional RL with quantile regression for learning the full return distribution. NAF ★ 28 Q-learning for continuous control using normalized advantage functions with PER and N-step returns. D4PG ★ 24 Distributed Distributional DDPG with an IQN critic, plus Munchausen RL and D2RL extensions. REDQ ★ 21 Randomized ensembled double Q-learning for sample-efficient continuous control. GARNE ★ 14 Genetic algorithm with recurrent networks and novelty-driven exploration for neuroevolution. GA Neural Network Optimization ★ 13 Genetic algorithms for neural network architecture search, hyperparameter tuning, and weight optimization. GANs ★ 12 ClusterGAN implementation — combining GANs with clustering for unsupervised representation learning. OFENet ★ 10 Online Feature Extractor Network — learned feature representations for continuous control RL. RA-PPO ★ 8 Risk-averse PPO — policy optimization with CVaR objectives for safer decision making. MBPO ★ 6 Model-Based Policy Optimization — using learned dynamics models to generate synthetic rollouts for SAC. PyTorch V-MPO ★ 5 V-MPO: on-policy maximum a posteriori policy optimization with learned temperature parameters. Hindsight Experience Replay ★ 4 HER for goal-conditioned RL — relabeling failed episodes with achieved goals to accelerate learning. PETS-MPC ★ 3 Probabilistic ensemble trajectory sampling with model predictive control for model-based RL. CEN Network ★ 2 Context-dependent Elastic Network for continual learning without catastrophic forgetting. D4PG-Ray ★ 2 Distributed D4PG using Ray for parallel data collection with IQN critic and Munchausen RL.