Candy Crush Saga from the famous gaming company King is a puzzle game with tons of levels. It's 2018 data, but it is said to add about 15 levels every week. (About 3700 per year)
I need to measure the difficulty of the level and balance it, but the problem is that it takes one week per level for human testers. To solve this problem, we created a virtual player that enjoys the game in a way similar to humans, and by collecting and learning the virtual player and human success rate data for the same level of data, it is said that the human success rate was predicted by the success rate of the virtual player. As a result, the time it takes to measure level 1 difficulty has been reduced from 1 week to minutes.
Of course, a virtual player is more like a reinforcement learning agent than a Human-Like AI, but the way to enjoy the game itself is not an efficiency priority like AlphaGo, but a “human style” that makes some mistakes like humans, I think it would be fun. Lift.
