強化學習與代理式 AI 的決策制定（結合 Python）培訓

Introduction to Reinforcement Learning and Agentic AI

• Decision-making under uncertainty and sequential planning strategies
• Core RL components: agents, environments, states, and reward signals
• The role of RL in fostering adaptive and agentic AI capabilities

Markov Decision Processes (MDPs)

• Formal definitions and key properties of MDPs
• Value functions, Bellman equations, and dynamic programming techniques
• Processes for policy evaluation, improvement, and iteration

Model-Free Reinforcement Learning

• Monte Carlo methods and Temporal-Difference (TD) learning
• Q-learning and SARSA algorithms
• Practical exercise: implementing tabular RL methods in Python

Deep Reinforcement Learning

• Integrating neural networks with RL for function approximation
• Deep Q-Networks (DQN) and experience replay mechanisms
• Actor-Critic architectures and policy gradient methods
• Practical exercise: training agents with DQN and PPO using Stable-Baselines3

Exploration Strategies and Reward Shaping

• Strategies for balancing exploration and exploitation (including ε-greedy, UCB, and entropy-based methods)
• Crafting reward functions and preventing unintended agent behaviors
• Techniques in reward shaping and curriculum learning

Advanced Topics in RL and Decision-Making

• Multi-agent reinforcement learning and cooperative strategies
• Hierarchical reinforcement learning and the options framework
• Offline RL and imitation learning for safer deployment scenarios

Simulation Environments and Evaluation

• Leveraging OpenAI Gym and creating custom environments
• Distinguishing between continuous and discrete action spaces
• Key metrics for assessing agent performance, stability, and sample efficiency

Integrating RL into Agentic AI Systems

• Fusing reasoning capabilities with RL in hybrid agent architectures
• Combining reinforcement learning with tool-using agent frameworks
• Operational considerations for scaling and deploying RL systems

Capstone Project

• Designing and building a reinforcement learning agent for a simulated task
• Analyzing training performance and fine-tuning hyperparameters
• Demonstrating adaptive behavior and decision-making within an agentic context

Summary and Next Steps