Learning to play games / Deep Reinforcement Learning
Submitted by Utkarsh Sinha (@liquidmetal) on Wednesday, 25 May 2016
Supervised and unsupervised learning techniques are relatively well understood in deep learning - Reinforcement learning is a new kind of learning that uses experience and interaction with the environment to make sense of the world around it.
An example of reinforcement learning is playing Atari video games. Creating a labeled dataset of “good moves” in the game is tedious and subjective. Unsupervised learning techniques might work but they don’t take advantage of the fact that the learning algorithm can interact with the game. Reinforcement learning combines the best of both techniques - no labels required and it can interact with the game world. This technique has also been used on the recent AlphaGo project by DeepMind.
- Brief demo and comparison to supervised and unsupervised learning [5min]
- Q-learning basics (dynamic programming) [10min]
- Deep Q-learning (gradient descent) [10min]
- Available DRL tools and quick start guide [5min]
- Q&A [5min]
- Techniques for faster training (if time permits)
Utkarsh Sinha a computer vision student at Carnegie Mellon University and currently at Microsoft Research. He’s been working in computer vision for the past few years and has been working in deep learning for the past year. His current research involves finding fine structures in images using deep learning techniques.
In the past, he has worked at DreamWorks Animation as a Technical Director and guided artists in modeling, texturing and lighting.