My agent receives $1, 0, -1$ rewards for winning, drawing, and losing the game, respectively. What would be the consequences of setting reward to $-1$ for draws? Would that encourage the agent to win more or will it have no effect at all? Is it appropriate to do so?
$\begingroup$ It can depend on many factors, for instance, your environment and data specifics. So, perhaps, there are no generic answers. Only some specifics observed during my experiments - it can slow down learning and create suboptimal agents which tend to make too risky decisions. Why don't you try it with your case and compare with my observations? That would be an answer for your question. $\endgroup$– Stepan NovikovNov 16, 2020 at 8:46
$\begingroup$ Yes, maybe you should tell us what zero-sum game you are trying to solve with RL, so that people could also try out your setting. $\endgroup$– nbroNov 16, 2020 at 12:47
$\begingroup$ Thanks for answers. This time I'm trying to solve the mancala game for different board sizes. My project is still on early stage (bugs/subtle bugs), so it's hard for me to really isolate the affect in this case. I will share the results if/when I make progress, but I can see now how it makes sense that this would experiment-specific. $\endgroup$– mark markNov 16, 2020 at 23:21