Loop to Learn

A Game of Reinforcement Learning

You are an engineer tasked with optimizing artificial intelligent agents in the completion of spatial navigation tasks to satellites. You have control over the time scale, which aspects of the environment triggers rewards in the agents, and when you want to select the best agent for testing (this is how you clear a level)--

The inputs the agent has available are 10 sensors that enables it to calculate distances of objects in front of it. The agent's output is to rotate and/or move forward. The performance (or fitness) of the best performing agent can be seen in the bottom right corner and is modulated by which factors rewards the agents in their actions.

For example, rewarding agents for the distance to the target means that the shorter the agents' distance is to the satellite at the end of the learning loop, the better its performance is. So be sure to decide how to reward the AIs or they might not even understand the task!

Every time 15 in-game seconds (can be scaled by you) go by, the agents are reset and the bottom half performing agents get updated neural networks through a mutation algorithm.