I am a bit confused regarding rewards in reinforcement learning. In my quite simple environment, where the agent has to find it's way to a target and kill it, the agent has control over heading direction, speed and firing. I have given the following rewards:

terminal states:

  • killing: 100
  • going outside boundary: -5

for each time step:

  • shaped reward f(dist, angle) which lies in [0,1] depending on distance to target and direction.

I am using PPO with Ray RLlib and somehow I don't get the desired behaviour. Sometimes the agent finds it's way to the target, but often not. My parameters are set as:

  • fully connected hidden layers: [128, 128, 64]
  • train batch size: 1000
  • time horizon: 100 (should satisfy)
  • learning rate: 0.0005

I don't know if I can to it like that, but I am wondering if my rewards make sense ?



You must log in to answer this question.