Checkpoints and new callback collection

• Add EvalCallback for proper evaluation using a separate env (--eval-freq, --eval-episodes)
• Add --save-freq argument to save periodic checkpoints of the model
• Add --load-best argument to enjoy script to load the best model according to evaluation
• Better seeding and fix possible race condition in folder creation
• Use new callback for hyperparameter tuning

Stable-Baselines v2.10.0 support and improvements

• Upgrade to Stable-Baselines v2.10.0

Many improvements:

• save on ctrl+c
• save training logs with the model
• specify a wrapper or list of wrapper in the config
• change hyperparams via the command line

New save format, ACKTR with continuous actions

• Upgrade to Stable-Baselines v2.8.0

Add Twin Delayed DDPG (TD3) trained agents

v1.3

TD3

HER support + Time wrappers

• Add support for Hindsight Experience Replay
• Upgrade Stable-Baselines version (v2.6.0) required
• Add wrappers for HER and for time constrained envs
• Changed default for hyperparameter tuning

Hyperparameter tuning + TRPO

• added hyperparameter tuning using optuna
• a2c for continuous actions
• upgrade stable-baselines (v2.5.1)
• add support for trpo + mpi training
• fixed frame stack loading

now more than 100 trained agents.

Soft Actor-Critic + Improvements

• Add Soft Actor-Critic (SAC)
• Tune bullet env for PPO2
• Improve logging

Video Recorder

Add video recorder