Note on MiniMax (Updating)
Motivation
This is a very traditional way of solving turn based game like chess or tic-tac-toc. It’s climax is Deep Blue AI in playing chess.
Note, some people think about GAN training procedure as a min-max game between G and D, which is also interesting.
So the setting is you have 2 parties with conflicting objective $L_A=-L_B$ . One side wants to minimize this value, the other side wants to maximize this value to win.
- In Chess it can be the value of all pawns summed, or a more complex evaluation of how a scenario favors black > white.
- In GAN, it’s the similarity or discriminability of the 2 distributions. G wants to minimize discriminability, D wants to maximize discriminability.