Neural Navigator

Neural Navigator is a fun little simulation where you can see a Machine Learning agent evolve to avoid obstacles.

The sim uses reinforcement learning to train the agent. There is a reward system to provide feedback to the agent for good/bad moves. The agent location is reset after every collision.

There is  no behavior programmed into the agent, but instead  emerges from the feedback. There is a positive reward given on every frame and a negative reward on collision.
The agent starts exploring randomly and then gradually exploits its learned knowledge.

You can find the reward details as well as the network weights in the Cora/Saved directory (reward_v_time.csv). There are 30 line traces as input to the neural network. There are two hidden layers with 32 and 16 neurons respectively. The weights are randomized every time the game is restarted. 

With these settings, it will take about 15 mins for noticeable learning and around 30 mins for a perfect run.

BGM: https://pixabay.com/sound-effects/wandering-6394/ 

RL Research: https://spinningup.openai.com/en/latest/