Release v0.16.0

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.

---

[v0.16.0] - 2026-03-17 16:13

Added

• Updated documentation and screenshots.
• More Simulation Settings in the TUI
  • RNN Tau for the RNNTrainer.
  • Discount/Gamma for the RNNTrainer and LinearTrainer.
  • Batch Size for the RNNTrainer and LinearTrainer.
  • Sequence Length for the ReplayMemory when the RNN model is selected.
• Metrics Class
  • Renamed the old HydraMetrics to HydraSnapshot and created a new HydraMetrics
  • When telemetry data arrives on the client, it is loaded into this object.
  • The class is used by the new GameScorePlot, LossPlot, ScoresDistPlot plots, and the HydraSnapshot class.

Changed

• The layout of the TUI.
• Complete Rewrite of Plotting
  • Replaced the monolithic TabbedPlot with GameScorePlot, LossPlot, and ScoresDistPlot.
  • The old TabbedPlot also managed the data used by the plots; this functionality has been moved into the HydraMetrics class.
  • The new architecture is cleaner, clearer, and easily extensible.

Known Issue

• Structural ReplayMemory Bug
  • The ReplayMemory that houses the training data needs a warm-up time to build a diverse training data set.
  • For performance reasons, the RNN's training data is stored in chunks that correspond to the RNN's sequence_length.
  • Previously, if an entire game's transitions did not make a complete sequence, then it was discarded. For a heavy RNN (large hidden layer size), which is slow to train, this resulted in a scenario where the replay memory never warmed up.
  • This will be addressed in the next release.