TAMER (Training an Agent Manually via Evaluative Reinforcement) is a framework for human-in-the-loop Reinforcement Learning, proposed by Knox + Stone in 2009.
This is an implementation of a TAMER agent, converted from a standard Q-learning agent using the steps provided by Knox here.
You need python 3.7+ with numpy, sklearn, pygame and gym.
Use run.py. You can fiddle with the config in the script.
In training, watch the agent play and press 'W' to give a positive reward and 'A' to give a negative. The agent's current action is displayed.
