Performance

  • Observation Time: 133.33

  • Reading action file IPC_recv.lua & take action: 2ms-5ms

  • Process State deltas: 20-50 ms

  • Time remaining for inference = 133 - 5 - 50 = 78 ms

To improve performance main work should be focusing on speeding up state parsing.

Ideas: * Move the stitching to the state process so more work is done in parallel

Inference

Inference will be parallel enough when using pytorch, the model will compute actions for all 10 bots actions at the same time (essentially batch = 10)