Braincraft challenge – 1000 neurons, 100 seconds, 10 runs, 2 choices, no reward (opens in new tab)

> How to submit? Make a pull request with your player, assumed performance and (short) description. I'll then re-run training and evaluation (with a random seed) and add the result to the leader board.

> Why 1000 neurons? Since It Takes Two Neurons To Ride a Bicycle, I decided that 1000 neurons should be more than enough to solve such a simple task.

[...]

  Bot: radius, speed, camera, camera.depths, camera.values, position, direction, energy, move_penalty, hit_penalty

  Environment: energy, probability, quality, leak, refill

Love it. The winning design is using handcrafted weights already.

The easiest thing to do is probably use A* to find the solution, then imitation learning on the NN to learn it. (The immediate feedback gets rid of the vanishing/exploding gradients problem).