Controller training random stop about world-models HOT 4 CLOSED

The VAE and MDRNN achieve decent results. The two errors you obtain are CMAES related. I guess this could be related to a n-samples set too low. In our replication, we used n-samples=16 and pop-size=4. If you could try with these values and keep me updated. I will investigate too.

from world-models.

Thanks so much! this fix worked! However, the training is going nowhere after 200 generations:
"python traincontroller.py --logdir exp_dir --n-samples 16
--pop-size 4 --target-return 950 --display --max-workers 12"

Loading VAE at epoch 33 with test loss 34.16893509338379
Loading MDRNN at epoch 29 with test loss 1.053613607351445
Loading Controller with reward 352.99061245415896
trainControllerGoingNowhereAfter200gen2.txt

from world-models.

Changed pop-size ==16 and
Current evaluation: -619.9075467084033 after 100 gens

from world-models.

it seems working

from world-models.