Gridworld constants and reward size fix

[casper2002casper]

Gridworld stores its constants as part of its environment instead of declaring them as constants, differing from other examples.
It also uses rewards in range [-10, 10] which creates two problems, first it overpowers the exploration term of the uct formula and secondly the value function of the SimpleNet network is limited by the tanh activation function to [-1, 1] so it's unable to output the required value functions.
Before

After

10% better reward for AZ, however the NN doesn't improve further as the max reward is reached after the second iteration.

[jonathan-laurent]

Thanks! Do you have any idea why the gap is wider between the neural network and AlphaZero in the second case?

[casper2002casper]

I think because the NN doesn't get replaced as the AZ benchmark maxes out

[jonathan-laurent]

Got it. This makes sense!