The authors describe a curious experiment held by several students from Russian University of Economics. They played several games of Go and could ask AI for a tip with the help of their coach. In total they tested four ways of using AI suggestions:
Option 1: The human player comes up with 3 candidate moves. AI gives them "traffic light grading": green, yellow and red. The human player chooses which move to play.
Option 2: The player determines an area of the board. AI shows the best move in this area.
Option 3: The player asks AI for a tip: where to play? AI shows currently the most important area of the board.
Option 4: AI gives a warning that there is a very important, possibly game-deciding move on the board.
It turned out that the optimal way of cooperation depends on the strength of a player. It's hard for a beginner to come up with 3 promising moves, they prefer being shown which area to play in. Higher level players might find it difficult to maintain their own strategy after hints and tips. They prefer being given suggestions at critical points, not at every move. More advanced players find it easier to use AI partners.
The first match of digital centaurs — pairs that consist of a human player and AI, will be held in Vladivostok during Genius Extreme festival. Virtual advisers will be developed by school and college students that are part of the National Technological Initiative.
The full article from "Dumai" will be published in our VK group.
