v0.23.3 - HOTFIX: Working Man 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.
[0.23.3] - 2026-03-29 @ 11:01 - HOTFIX - Working Man Release
This release fixes the performance collapse from the previous releases.
Added
- Enhanced Snapshot Report
- Increased the granularity of the memory and epsilon nice tables from every 500 to every 100 episodes.
- Added the current gear to the memory bucket table.
- Added the recent mean and median to the mean and median table.
- Created sane baseline configurations for the Linear, RNN, and GRU that avoid destructive training behaviour and promote better recovery behaviour under stagnation and hard reset conditions.
Fixed
- Modify the
ATHDataStoreto return the complete list of buckets, even if some of them are empty. - Complete overhaul of the ATH Memory modules to ensure they are correct.
- Added a LOT of asserts and data validation steps.
- Basically improved the stability and safety of the Replay Memory.
EpsilonNicewasn't interacting correctly with theGameLogicfor collision detection- Refactored the
GameLogic:step()function to isolate the collision detection functions from the reward calculations - Updated the
GameLogic:would_collide()method (used byEpsilonNice) to use the refactoredGameLogic:step()changes.
- Refactored the
- Fixed Multiple Config Propagation Bugs
- Ensured that the stagnation threshold and critical stagnation threshold TUI settings made their way into the
TrainMgrmodule. - Ensured that the Epsilon Nice TUI settings made their way to the
EpsilonNicePolicymodule.
- Ensured that the stagnation threshold and critical stagnation threshold TUI settings made their way into the