This workshop is ideal for those who have a familiarity with deep learning and basic maths who want to break into the game changing world of Deep Reinforcement Learning (DRL) whether those who have a passing curiosity in this field to those who aspire to use this to develop autonomous agents to digital advertisement to machine translation.

This workshop starts of with a brief introduction to Markov decision processes and reinforcement learning. (RL) We will look at what separates reinforcement learning from supervised and unsupervised problems and examples that fall under the category of RL. We will investigate two class of methods to solve these: Value based methods and policy based methods.

One of the exciting challenges that policy-based methods in Deep Reinforcement Learning (DRL) conquer is calculating the intractable loss function and its gradient which we resolve using the log derivative trick.

The hands-on part of the workshop involves using Tensorflow to construct and train two DRL architectures: Deep Q-learning and the REINFORCE algorithm. We will use the OpenAI gym library which provides a great API for training and visualising our RL agents in game environments. You will design and train your own agents to quickly master classic games like Pong and Pinball without providing them any information about what moves are good or bad.

We will wrap-up the workshop by looking at some recent applications and developments in DRL and some of the ongoing challenges in this field.

The ideas explored in this workshop are transferable skills to other applications in deep learning and provide a great stepping stone for those that have a good familiarity with deep learning in applications such as image classification and want to explore more advanced topics in deep learning.