Deep RL - Achieving Superhuman Performance at Games
The Syzygian Inferno (~Syzygianinfern0) |
Have you ever speculated how computers would accomplish tasks such as playing games? Thanks to the current state of processing power and learning algorithms, computers can achieve lightning fast responses, develop strategies and beat the games with educated moves.
While neural networks are responsible for recent breakthroughs in problems like computer vision, machine translation and time series prediction – they also associate with reinforcement learning algorithms to produce something astounding like AlphaGo.
Reinforcement learning refers to goal-oriented algorithms, which acquire knowledge about attaining a complex objective (goal) or maximize along a particular dimension over many steps; for example, maximizing the points won in a game. They can start from a blank slate, and soon supersede us to achieve superhuman performance. Like a child incentivized by spankings and candy, these algorithms are penalized when they make the wrong decisions and rewarded when they make the right ones – this is reinforcement.
The workshop intends to acquaint one with the process of creating personal game environments and formulating them into Markov Decision Processes and successfully train RL agents to master it.
We will start with classic environments under gym wrappers from OpenAI to introduce participants to underlying algorithms for RL agents. Action based algorithms involving Q Learning will be taught. Deep Learning using TensorFlow will be the method of incorporating the agent’s brain. Original research papers will be quoted and further ideas and concepts to explore will be discussed as by the participants will.
Then proceed into training RL agents to play in intermediate 2D environments such as Atari and PyGame environments such as Chrome T-Rex game. This is to familiarize the participants on the concepts of training RL agents based upon image data in the form of frames and processing them using Convolutional Neural Networks. Insights into research based on famous projects such as AlphaGo and other marvels of the RL community will be discussed. Deep learning will be done using TensorFlow, although participants may resort to their favoured frameworks for these purposes.
Participants will finally accommodate these algorithms into 3D environments using Unity ML Agents Toolkit. Policy based methods such as Policy Gradients and Actor Critic will be utilized to counter advanced states and actions. The current state of Reinforcement Learning community and latest research papers, advancements, competitions that participants can explore will also be provided.
- Knowledge of High School Mathematics
- Deep Learning Basics
- Experience with TensorFlow, PyTorch or any other DL framework
- Laptop with Internet Connectivity
The workshop will be conducted by S P Sharan and Sachin Kumar, students at the National Institute of Technology, Tiruchirappalli. We are highly enthusiastic young minds aiming to contribute to the Machine Learning Community.