3 October 2019
We will do something different this time, a workshop, not a sprint.
By engaging the revolution of AI and deep learning, reinforcement learning also evolve from being able to solve simple game puzzles to beating human records in Atari games. It also opens the possibility of using reinforcement learning in making real life decisions. It’s time to do some serious DRL. (Or just play some games)
In this 3 hours workshop we would introduce some deep reinforcement learning (DRL) algorithms, as an exercise we will implementing them in python with deep learning libraries, specifically keras and tensorflow, to play games in Open AI Gym and simulated Atari.
In the first section, we will touch on the basic in reinforcement learning and implement using crossentropy method to play a simple games, on top of implement the basic tabular crossentropy method, we will also implement deep crossentropy method when the table which keep track of the policy becomes too large.
Most of the problem in the real world are model-free setting, i.e. we don’t know what the final result will be like for our intermediate actions. In the second section, we will introduce Q-learning and SARSA, two model-free policies which involve understanding of Bellman equations. We will also start introducing experience replay buffer which is essential to speed up learning.
In the last section, we will explore using DQN (Deep Q-Network), which is a network develop by Google Deep Mind involve using CNN as an agent to play Atari game. Experience replay buffer will also be implemented to speed up learning.
As the end of the workshop, participants should be able to understand the concept of the deep reinforcement learning algorithms that we covered, implement them in python with keras and tensorflow.