andri27-ts/Reinforcement-Learning

Features

• Reinforcement Learning Curricula - Provides a structured curriculum of lectures and Python code for learning deep reinforcement learning.
• Actor-Critic Architectures - Implements actor-critic architectures that combine policy-learning actors and value-estimating critics to improve training convergence.
• Evolutionary Strategy Optimization - Includes a library of evolutionary strategies and genetic algorithms as alternatives to gradient-based learning.
• Deep Q-Learning Implementations - Provides deep Q-learning implementations using neural networks and experience replay to master game environments.
• Deep Reinforcement Learning Implementations - Provides deep reinforcement learning implementations for solving control tasks using value-based and policy-gradient methods.
• Evolutionary Strategy Libraries - Ships a dedicated library of evolutionary strategies and genetic algorithms for agent optimization.
• Policy Gradient Methods - Develops agents for complex control tasks using gradient-based policy optimization and actor-critic architectures.
• Proximal Policy Optimization - Provides a surrogate objective function optimizer to ensure stable policy updates in continuous action spaces.
• Model-Based Planning - Features a model-based system for predicting environment states and rewards to enable internal simulation and offline planning.
• Model-Based RL Systems - Features a model-based RL system that enables internal simulation and offline planning.
• Policy Gradient Optimizers - Uses gradient-based optimization methods to directly adjust agent behavior within actor-critic architectures.
• Deep Value-Based Training - Implements value-based learning techniques including deep Q-networks and SARSA variants for discrete environments.
• State-Action Value Updates - Implements state-action value updates by calculating errors between target rewards and current value predictions.
• Double DQN Implementations - Implements Double Deep Q-Networks to decouple action selection from value estimation and reduce overestimation bias.
• Discrete Environment Solvers - Trains agents to find optimal action-value functions specifically for discrete environment settings.
• Environment Dynamics Modeling - A model-based approach that builds internal representations of the environment to forecast future states and optimize agent actions.
• Experience Replay Buffers - Provides experience replay buffers to store and sample past interactions, breaking data correlation for stable training.
• Exploration Strategies - Provides greedy and epsilon-greedy action selection strategies to balance exploration and exploitation.
• Noise-Based Exploration - Implements noise layers within the network to manage the exploration-exploitation trade-off without epsilon-greedy methods.
• Genetic Algorithms - Employs genetic algorithms to optimize agent parameters as a scalable alternative to traditional gradient-based methods.
• Clipped Policy Objectives - Uses a clipped surrogate objective function to ensure stable behavior updates in continuous action spaces.
• Neural Dynamics Models - Implements neural dynamics models to predict future states and rewards, enabling planning and knowledge transfer.
• Environment Simulation Planning - Builds internal neural network models to predict future states and rewards for action optimization within simulations.
• Dueling Network Architectures - Implements neural network architectures that decouple state value estimation from action advantage for improved stability.
• Sarsa Update Implementations - Implements the Sarsa on-policy temporal difference learning algorithm using deep neural networks.
• Multi-Step Return Calculations - Provides a forward-view multi-step approach for updating target values to increase learning efficiency.
• Multi-Step Returns - Implements multi-step temporal difference learning to propagate value information faster through the state space.

Star history