how to improve the convergence performance of training loss?

[williamyuanv0]

Hi kengz, I find that the convergence performance of training loss (=value loss+policy loss) of ppo algorithem applied in game pong is poor (see Fig.1), but the corresponding mean_returns shows a good upward trend and reaches convergence (see Fig.2). That is why? how to improve the convergence performance of training loss? I tried many imporved tricks with ppo, but none of them worked. Fig.1 Fig.2