Online MLMU #9b: Reinforcement learning 1: deep Q networks – Michal Chovanec

Language of the talk: Slovak

The meetup will be hosted online - using Google Meet platform.

Syllabus of the talk:
1. Reinforcement learning versus supervised learning
2. Deep Q Network (DQN) – why it’s naive NN not working

3. Advanced architectures: noisy DQN, duelling DQN, rainbow DQN, attention DQN and hands-on practices how to create a network architecture

4. Real examples (Atari, Super Mario, Doom 2) – how is network interacting with the given world.

The whole talk is full of real examples of PyTorch integration and results. Reinforcement learning is a machine learning method, which is trying to find an optimal strategy (trajectory) in Markov decision process (in this case, game). Agent (usually deep neural network) is actively interacting with the environment, considering current status to choose actions (policy) as output, evaluating rewards and punishments. The goal is to maximize the sum of rewards. The agent does not know what is right behaviour but has knowledge of actions and rewards. Based on this, reinforcement learning is capable of finding a better strategy than human.

Michal Chovanec is a researcher in AI field. Currently is working in Tachyum in Bratislava, developing AI accelerator of the new processor. His PhD degree is from University in Žilina. His favourite topics are reinforcement learning, robotics (winner of Gold Medal in Istrobot challenge) and modelling of red blood cell in research group Cell in fluid (FRI, UNIZA). His hobbies are hiking, archery and martial arts.