Hello,
In my project, I am doing multi-agent reinforcement learning for motion planning application with obstacle avoidance, using GNNs.
In my graph structure, I define an edge name (obstacle_to_agent) which uses distance as an edge feature. To mimic real world scenario, I assign an edge between an obstacle and an agent if their distance is within the agent’s sensor range.
I am wondering if an agent is far away from an obstacle, then should I assign an edge between the agent and a dummy obstacle with an edge value of 0? Or, should I have no edge assignment between that particular agent? I am curious on how does it impact training and performance.
Thanks