Home

Awesome

Results of AdaBelief and Adam on Reinforcement Learning with SAC (Soft Actor Critic)

Dependencies

How to run

sh run_adabelief_walker2d.sh sh run_adam_walkerd.sh

You can change --env Walker2d-v2 to --env HalfCheetah-v2 for different tasks

Hyper-parameters

eps for AdaBelief is 1e-12, other parameters are default as in adabelief-pytorch==0.1.0

Optimizerlrbeta1beta2epsilonweight_decayweight_decouplerectifyfixed_decayamsgrad
Adam1e-30.90.9991e-80.0----
AdaBelief1e-30.90.9991e-120.0True=FalseTrueFalseFalse

Results

<img src="HalfCheetach.png" width=750> <img src="walker2d.png" width=750>