Ax3l uses Snake Lab to train a Transformer with four attention heads for 750
episodes. The LLM chooses hidden width from 32 to 128 in steps of 8, encoder
depth from 1 to 3 layers, and dropout from 0 to 0.1 in steps of 0.01. The
defaults are width 64, one encoder layer, and dropout 0.05. The target update
rate (tau) is fixed at 0.05.
Actions are left, straight, and right relative to the current heading.
Its 24-value feature vector contains:
There is no full-board view or food-distance measurement. Sequence length controls how many consecutive frames training uses. Hidden size controls Transformer width; batch size counts sampled sequences. Learning rate controls update size, and gamma discounts future rewards. Choice rules are in your submission tool.
The move limit is max_moves_multiplier times the current snake length.