v0.27.0 - Fisheye Release

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[0.27.0] - 2026-04-10 @ 07:38 - Fish Eye Release

Changed

Fish Eye - Achieves a high score of 80 with the linear model

• Wrote a new GameBoard:get_state() function that transforms the game board into an egocentric frame of reference, giving the agent a consistent “forward-facing” view of its surroundings. Combines dense local spatial encoding (7×7 grid) with abstract directional signals to guide decision-making.
  • Rotates the world so the snake never has to learn “up/down/left/right” separately
  • Gives it a local radar scan (7×7 grid)
  • Adds a compass ping toward the food
    Cooldown Reset Feature
• At the end of each episode, if a new high score has been achieved:
  • The TrainMgr signals the ATH Gearbox
  • The gearbox resets the cooldown counter, giving the NN a chance to continue to mine the training data using the current sequence length and batch size.
• Updated PyPI and RTD documentation

Removed

• Monte Carlo Tree Search: It's a terrible, computationally expensive way to play the game.

Fixed

• On reset, the initial epsilon widget was being passed to the HydraMetrics class instead of the actual initial value of epsilon.