强化学习 - Reinforcement Learning

课程介绍 Course Introduction

学分：3 | 先修课：机器学习导论、概率论与数理统计 | 学期：第6学期

强化学习研究智能体通过与环境交互、试错优化策略以最大化累积回报的学习范式。课程内容包括马尔可夫决策过程、值迭代与策略迭代、蒙特卡洛方法、时序差分学习（Q-Learning、SARSA）、Deep Q-Network、策略梯度方法（REINFORCE、PPO）、Actor-Critic框架及多智能体强化学习。学生将掌握序列决策的核心算法，应用于游戏AI、机器人控制、推荐系统等场景。

Reinforcement Learning studies how agents learn optimal policies through environment interaction and trial-and-error to maximize cumulative reward. Topics include Markov decision processes, value and policy iteration, Monte Carlo methods, temporal difference learning (Q-Learning, SARSA), Deep Q-Network, policy gradient methods (REINFORCE, PPO), Actor-Critic frameworks, and multi-agent RL. Students master sequential decision-making algorithms applicable to game AI, robotic control, and recommender systems.

大作业 Final Project

作业标题：Atari游戏AI智能体

学生需基于DQN或PPO算法训练一个Atari游戏（如Breakout、Pong）智能体。要求使用Stable Baselines3或自行实现算法，完成环境搭建、训练、调参与评估，平均回报需超过基准线。需提交代码、训练曲线、性能对比报告及演示视频，讨论探索策略、奖励塑形与稳定训练技巧的影响。

Students train an Atari game agent (such as Breakout or Pong) using DQN or PPO algorithms. The project requires setting up the environment, training, hyperparameter tuning, and evaluation using Stable Baselines3 or a custom implementation, with average reward exceeding the baseline. Deliverables include code, training curves, a performance comparison report, and a demo video, discussing exploration strategies, reward shaping, and stable training techniques.