Add previous model action (so if a lookup parser chooses action 89, you have it)

[Kaiyotech]

I have no idea if this is valuable enough to anyone except me to be worth it. It's currently a breaking change, but perhaps someone smarter than me can take a look and see if it's possible to do it better, or just if it's worth making it breaking for rlgym 2.0 maybe.

[lucas-emery]

What's the difference between previous_action and previous_model_action? I see the reward fn is getting both

[lucas-emery]

Ah I see, it's the action before being parsed

[lucas-emery]

How are you using this value in your reward fn and obs builder? I'd like to understand the usecase a bit better

[Kaiyotech]

In the rewards I punish switching actions to encourage sticking with one submodel (this was your idea, lol). I think that may or may not have value for a normal lookup parser model.

In the obs I normalize it (divide by number of lookup possibilities so it's from 0 to 1) and then put it on an action stack. I can have a 30 stack of previous actions (a full second at 4 tick skip) in only 30 obs spaces instead of taking 240 spaces like normal. I think that part might have value for a lookup parser.

On Fri, Dec 23, 2022, 6:38 AM lucas-emery @.***> wrote:

How are you using this value in your reward fn and obs builder? I'd like to understand the usecase a bit better

— Reply to this email directly, view it on GitHub https://github.com/lucas-emery/rocket-league-gym/pull/54#issuecomment-1363879398, or unsubscribe https://github.com/notifications/unsubscribe-auth/AWLB4KV77PYCF4NDJUKPVPLWOWFLPANCNFSM6AAAAAASV4OOXE . You are receiving this because you authored the thread.Message ID: @.***>