instadeepai / Mava

🦁 A research-friendly codebase for fast experimentation of multi-agent reinforcement learning in JAX
Apache License 2.0
709 stars 83 forks source link

Feat/learning rate decay #998

Closed RuanJohn closed 8 months ago

RuanJohn commented 8 months ago

What?

Add the option to linearly decay the actor and critic learning rates during training.

Why?

Based PPO implementation details blog here.

How?

Simple utils file that creates a linear learning rate decay scheduler and passes that to the optimisers.

Extra

Default behaviour (constant learning rates) are retained by setting decay_learning_rates: False in the system configs.