AI plays Trackmania - Map5  

The agent tries to predict the remaining racing time if it executes action A, B or C. The impact of wiggling on overall racing time is probably so small that the agent is unable to differentiate the actions, as long as they accelerate. There is no reward for the agent to press as few buttons as possible, or for it to keep the same action several frames in a row.

@@linesight-rl this also can be caused by stochastic policy, on evaluation it's better to disable all randomness

@@linesight-rl cool project anyway!

@@howuhh8960 In this case, there is no stochastic policy. The reinforcement learning algorithm is value-based.

@@linesight-rl very cool, keep going!