Automatically create actors for incoming requests

RLlib
1

How severe does this issue affect your experience of using Ray?

  • None: Just asking a question out of curiosity

Hi,

I am wanting to make a system that has a trainer (PPO) fully set up locally on a server, then the client would send requests for predictions from an environment. The server would then create an actor (or process?) that creates a new ID for the session and essentially trains on the incoming requests.

The idea here is to create a game and have clients make HTTP requests for predictions (the latency does not affect the gameplay). I am aware of Unity ML agents, but the clients may be connected/disconnected at will and from environments that don’t have python installed for example.

Is this possible? If so, can anyone point me in the right direction?

Thanks!

Edit:
In particular, there is the policy_client + policy_server. Can I have a request be sent to the policy_client whenever I need a new game (from a client) which would use the new method to startup a new game that is running parallel to the other games? policy_client then forwards all the requests as normal to police_server.

So instead of policy_client running the game, it is getting requests with all the information from different devices, and acting as a mediary for policy_server

2

Hi @Denys_Ashikhin ,

Have a look at ExternalEnv and PolicyClient.
I can’t point you into the right direction though, you’ll have to code this up yourself.
You can modify PolicyClient to your liking.