A framework that allows for implementing various reinforcement learning models that I created so that I could:

• Learn RL
• Understand each implementation
• Compare their performance to deepen my understanding
• Deepen my understanding of tensorflow / keras

This will be done with an environment (created by me) that plays Hanabi, but should be generalizable to any environment.

Currently Q-learning is implemented and benefits from GPU acceleration. I get roughly 25k samples trained / second on my desktop, and 10k / sec on my laptop.

One can create an instance by running main.py and then the following in the console:

instance = wrapper()

and train with

instance.train(10)

• Plots steps per episode and average reward on one graph every plot_frequency
• Model saving and loading
• Performance tracking (how long the wrapper is spending performing various tasks, both as a percentage and absolute value)

• Double DQN (versions 1 and 2 or disabled)
• Discrete agents (whether experiences are stored per player or as a single player)
• Customizable discount
• Customizable alpha (how much towards a new value each training batch should update)
• Customizable optimizer and learning rate
• "Greedy" and "Epsilon Greedy" Bandit algorithms
• Different numbers of players (2-5)
• Customizable hidden layers for model prediction
• Different batch size
• Customizable regularization (l1 only)
• Customizable memory size

See the comments for the options under wrapper for a description of available options.

• Multiple threads for data generation and training
• Dueling DQN