With the wide use of machine learning, people have higher expectation of such a technique, and wish to use machine learning to deal with decision making tasks, which are more challenging than prediction tasks. Decision making tasks are actually quite common in our daily life. Examples include, to name a few, game AI, autonomous driving, dialogue robots, interactive recommender systems, intelligent transportation light control etc. Compare with the prediction tasks, the most appealing difference of decision making tasks is the output of the agent, i.e., the action, will delivered to the environment and change it, which requires the agent to perform multi-step or even longer horizon planning. Such a learning-from-interaction and long-term planning paradigm is called reinforcement learning.

This course provides a comprehensive introduction of reinforcement learning techniques, including the fundamental concepts and math of reinforcement learning, and basic & advanced methodologies like MDPs, dynamic programming, temporal difference learning, value function learning, model-free control, policy gradient, actor critic, deep reinforcement learning, imitation learning, multi-agent reinforcement learning etc. Additionally, the coursework includes hands-on tasks, in which the students are required to design machine learning programs to accomplish several intelligence tasks, and are high encouraged to further improve the reinforcement learning agent performance via trying different models and upgrading the code implementation.