About the implementation
albertcity opened this issue · comments
Dear Authors,
Does there exist any implementation of Self-Supervised Attention-Aware Reinforcement Learning?
Besides, I am curious about how to set the hyperparameters in Eq.(2) and Eq.(3) (i.e.
Many thanks!
:)
Hi, @albertcity
The code has been uploaded. The sparsity penalty (1e-3) and reconstruction threshold (1e-2) were set during initial single-task experiments on MsPacmanNoFrameSkip-v4, and then used the same value for all other envs shown in the paper.
After paper published, a sweep experiments were conducted to find the best value of those two hyper-parameters for more Atari envs, and the results show the best value can vary from env to env.