Yomi Layers

Knowing the Mind of the Opponent

David Sirlin wrote a series of articles on playing to win called, unsurprisingly, "Playing to Win". Yomi Layer 3: Knowing the Mind of the Opponent talks about predicting what the opponent will do in situations where the opponent is aware they they are being predicted. One potential option for Technical Machine, in addition to depth of search, could be the yomi layer. Yomi level 0 for Technical Machine would be how it works now: examine all moves, and make the move that gives the foe the least potential for gain. In other words, Technical Machine always plays the "safest" move.

Yomi layer 1 AI would be for the AI to recognize that the foe may be using a similar algorithm to Technical Machine. Therefore, it may generate stronger play to apply Technical Machine's algorithms to the foe's move selection process and determine what their "safest" move is. Rather than trying to stop their highest payout move, which may be so risky that the opponent never uses it, Technical Machine could attempt to counter their "safest" move instead. In other words, assume that the foe thinks like the AI does, and play the best counter to that move instead. Of course, TM would still have to play its own safest move some times, to prevent the foe from adopting this same strategy to gain against TM.

A definition of higher yomi layers would be to take the original algorithm to determine TM's safest move, and then assume that the foe knows what this move is. They know exactly what move Technical Machine would want to use. Therefore, they also know what move would best counter that move. Technical Machine could fully evaluate the response to their counter to TM's best move, and find the move that scores best assuming they use that counter. This would be the counter to an opponent thinking at yomi layer 1, as explained above.

This strategy could be implemented as a mixed strategy by estimating how likely it is that the foe is at each yomi layer in their own play. It then plays at each yomi layer + 1 with that probability. The downside of this is that it reduces the effectiveness of alpha-beta pruning, because it means that TM has to fully evaluate more moves. However, it could theoretically improve play, and is much faster than searching one more move ahead.

Technical Machine does not currently implement this strategy.