I've been having issue with getting my DQN to converge to a good solution for snake. Regardless of the different types of reward functions I've tried, it seems that the snake is indefinitely going around in circles. I have not tried exploring more states yet because I am confused about how to properly assign reward.

Currently, I am using a 2D-Gaussian distribution to assign reward where $f(x=x_{food},y=y_{food}) = 1$. Terminal states like hitting the wall or itself result in a reward value of -1.

My reason for using the Gaussian was because of the relatively sparse rewards in this game and the ability to easily clip rewards between [1,-1] in meaning full way.

I have two questions.

  1. Is this an appropriate way to define the reward function?
  2. Currently during training, I do no accumulate the reward during each play iteration. So each reward for transitions are independent of the reward values before it. Right now I am doing $\big[S_1,A_1,R_1,S_2,A_2,R_2,S_3\big]$. I've looked at other code where people have accumulate the reward like $\big[S_1,A_1,R_1,S_2,A_2,R_2 = R_1+r,S_3\big]$. Where $r$ is given by the reward function. The thing is, I can't find a paper that defines why you should do this. So my question is, which way is the appropriate way to assign reward?

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.