README.md

Q-learning

Using reinforcement learning to play snake and similar games.

An environment is created that simulates a game . This environment takes actions and returns the resulting screen of those actions, plus the resulting reward (-1 if the player dies, 0 if nothing happens and 1 if the player scores).

A player with a neural-network provides actions and learns from the environment responses using Q-learning per advantage learning. The network should learn what actions provide the best value.

Techniques used

• Advantage learning, particularized to Q-learning.
• Double Q-learning
• (Prioritized) memory replay
• Progressive discount rate growth
• Progressive exploration rate growth

Results

For such a simple game the player should be able to learn to play for much longer, but it is clearly working:

More info on Q-learning

Demystifying Deep Reinforcement Learning