v0.2.0 - 2025-04-22
🚀 Features
- (callbacks) Enhanced callbacks for flexibility.
- (force) Added
force flag to save methods for file overwrites.
- (ppo) Added
LiquidPPO algorithm.
- (handler) Added file saving for training completion details.
- (ncp) Added multiple weight initialization options.
- (sac) Added
LiquidSAC agent for continuous action spaces.
- (sac) Added
LiquidSACDiscrete agent for discrete action spaces.
- (neuroflow) Added main logic for
NeuroFlow.
- (agent) Added
NeuroFlowDiscrete agent.
🐛 Bug Fixes
- (ddpg) Fixed noise handling and prediction bugs.
- (cell) Fixed
sparsity_mask assignment bug.
- (params) Fixed parameter counts in in DDPG.
- (ppo) Fixed PPO callback bugs and metric tracking.
- (config) Fixed bug with
train_params in RLAgentConfig.
- (load) Fixed model loading bug.
- (buffer) Fixed
warm method bug when num_envs=1.
- (buffer) Fixed save bug where directories don't exist.
💼 Other
- (box) Added Gymnasium box2d environments by default.
🚜 Refactor
- (ncp) Added
update_mask helper methods.
- (metrics) Updated training metrics name for clarity.
- (metrics) Simplified metric classes using base class.
- (train) Refactored
TrainHandler, TrainConfig to simplify.
- (buffer) Added Actor hidden state to buffer.
- (seed) Improved random seed generation.
- (save) Simplified
save, load method implementations.
- (sac) Moved
SAC agents to separate folder for simplicity.
- (ncp) Renamed
NCPModule -> LiquidNCPModule for clarity.
- (agents) Refactored framework to centre around
NeuroFlow.
- (save) Moved
completed.json to save directory.
- (warm) Improved buffer warming step implementation.
- (utils) Simplified
capture utility methods.