Open 4SkyNet opened 7 years ago
Test new A3C-LSTM
with 8
threads as underlaid based for perception and worker (gym's Pong
environment)
Test (one more time) vanilla A3C-LSTM
with 8
threads & 20
steps (last gym's PongDeterministic-v4
)
315 steps/sec
on c4.xlarge
with only 4
vCPU
PS> gradient clipping added to compare with previous one (resized one RGB frame as input)
TODO: Distributed version of FuN - FeUdal Networks for Hierarchical Reinforcement Learning (original paper)