Strange behaviour when team is full about super-ml-pets

[andreped]

I think it is easier if you just test the model during deployment. You should see some bugs quite early on - especially as the team gets full. I'm not really sure what is causing it, but take a look.

I would recommend having two monitors, one displaying the game where the AI is playing, and the other showing the terminal.

[andreped]

This might be related to this issue: #15

[GoldExplosion]

I think the Pet is not sold but pilled(you probably didn't catch it since it was happening very fast). There is an issue with pilling a pet in this framework probably because of the following piece of code

super-ml-pets/src/agent.py

Line 74 in a5f7be6

if env.player.lives <= 3:

It is not a problem with sapai itself as pilling works properly in the engine's tests. Pilling a Pet adds -1000 to the Pet's health. The above code keep the AI player from dying but also breaks the sapai engine's pilling mechanism and allows Pet with -1000 health to continue playing.
We need to know if the AI won or lost against real opponent so that we can avoid using altering the number of lives remaining.

[andreped]

Makes sense. I think it is very smart of us to get the deployment stuff working properly, before doing a lot more on the RL-aspect. I believe the model I have trained is actually decent, it just fails to win more than 2-3 games, as bugs seem to occur when the team is full (or after shuffling a lot). Not really sure what the main reason is.

Would be great if you could take a look at it.

[GoldExplosion]

Yes, will do.

[andreped]

The issue I observed was not related to the team being full, but rather to how teams were reordered. This has now been fixed in fa8516b.

Could be that there is a strange behaviour if an animal is sold, but I have not observed it yet. Could create a separate issue if the sold event is broken.