Flat A3C agent

[dai-dao]

Hi,

I really like your work, and want to ask for some clarifications on your new observation on training a flat A3C agent without the meta-controller. In this case are the sub-goals randomly generated every 'c' timesteps? (instead of the meta-controller outputting the sub-goal)

Thanks,
Dai

[Nat-D]

Hi Dai,

Yes, the sub-goals were randomly generated every c=100 time-steps. I also found that fixing the sub-goal to be just the first one also works in some seed. This only works with feature-control pseudo reward tho.

Best,
Nat

[nina124]

Does "randomly generated meta-action" work with pixel-control pseudo reward?