Closed glistering96 closed 1 year ago
I've found a severe bug when using Gymnasium's SyncVecEnv, in which the episode is not terminated. #3 is wrong and slow compared to the numpy vec environment. So I just remove the gymnasium Sync Vec Env.
Note that numpy TSP env runs around x2 faster than the SyncVecEnv
To boost up the training, numpy vector based environment seems to be better.
I have implemented it on the branch np_vec_env with a test bed.
Will merge into the master if it is validated to work ok.