Current progress & Near‐term plan

Current progress

• This version uses CNN based representation network + LSTM based dynamics network architecture for RGB input environment (LunarLander-v2, using rgb states wrapped by PixelObservationWrapper(self.env))
• It works for randomly initialized env(random terrain and acceleration), but still not perfectly converge to more than 200 score.
• Training(non-fixed env seed) takes several hours, longer experiment length than 5000, with Deep CNN.
• Training(fixed env seed) takes up to 1 hour, experiment length 200~2000, with Deep CNN. Agent can learn perfectly in this case.
• It needs to be improved further.

Near-term plan

• Write the docs on the wiki.
• Check the nni config setting and try to fix the issue#5.
• Efficiency optimization about code and nni experiment setting.
• Try to use off-policy correction method like V-trace, Retrace.

Previous

Progress

• contents