koobas.hobune.stream
ML Agents learn to stop trying to achieve their goal and instead sit and do nothing.
I'm trying to learn how to use ML Agents. I tried to make a simple game where the ML Agents just have to navigate to a goal and avoid the red area, which would kill them. However, the agents eventually decided that getting to the goal was too much work and that the best option is to sit there and do nothing. I first gave them a reward for reaching the goal and a negative reward for hitting the red. They weren't going far enough to get to the goal, so I gave them a very small reward for going further towards the goal.
At first, everything was going normally. They started off touching the red a lot but eventually got a bit closer to the goal. Then, a few of them started to reach the goal. They started getting closer to the goal and more of them were reaching it, and then I left it go for a while. About a half an hour later, I checked it and saw that they stopped moving. I checked and it was still running and they were still training, but they had stopped moving.
How can I encourage them to achieve their goal? Should I give them a bigger reward for completing it? A bigger negative reward for them dying? Change the positive reward for being closer to the goal to a negative reward for being farther? What should I do? My current rewards are 1 for reaching the goal, -0.5 for dying, and 0.001f times their z coordinate (the goal is around z 60)
