Performance when multiple agents exist

Hi,

I am experimenting to see how it performs when there are multiple agents in a game.

Currently, there are two agents, both of which act with the same NN structure model. The processing time and the behavior on the profiler are different when both are given a model with the same weights and when each is given a model with a different weights (see the attached images). I am looking at the AcademyFixedUpdateStepper.FixedUpdate function as the agent processing time. Also, the times taken by the root.DecideAction function and the root.AgentAct function, in the Academy, are very different between the two cases.


What do DecideAction and AgentAct functions do respectively? Also, How does it work when there are multiple agents (with same network and with different networks)? How do they change when the number of agents increases?

Thanks

 

Does anyone have any idea on this? I appreciate if if you could provide any pointers.

 

I know its old thread.

But I've the same issue with 8 agents 4vs4 game, root.DecideAction take too much time to compute, and since the game is webgl its even slower since everything is single threaded.

Yet, one from unity dev team suggested that num_layers & hidden_units paramaters for the training configuration playing strong part effecting the performance for decision making

So less num_layers and less hidden_units the better performance it would be.