Gradients always explode, who will help me

[chamovalera]

at the beginning of the training this always happens as possible, initialize my weight, decrease the cup, modify the weight but as

you want the algorithm explodes

[ptrblck]

I don’t know what PC stands for, but assuming your loss explodes you could check if e.g. the learning rate is too high and reduce it if needed.

[chamovalera]

[chamovalera]

I’m using DDPG TD3 That’s the actor’s loss, and the low critic’s is normal:

[J_Johnson]

Your charts are not clearly marked. Can you include labels on the axes? (or write them in, i.e. x-axis: time per game, y-axis: n-game)