An AI can beat a chess grandmaster. An AI can become the StarCraft esports champion. But creating an AI that could play Pokémon at the competitive level has been a more elusive problem.
Thanks to the variety of monsters, stats, moves, and items, a Pokémon battle has hundreds of thousands of factors for any player — or machine — to consider. But that hasn’t stopped some people from trying. Most recently, Future Sight AI, created by computer scientist Albert III, successfully made it into the top 5 percent of the competitive ladder.
Albert posted a video explaining how it all works, but to summarize, the bot takes in all the information it can about the current state of the game, extrapolates the possibilities for all the turns it could take, looks a couple of turns ahead to how these would play out, and then chooses the option that can lead to the highest number of best outcomes. By doing all of that within 15 seconds, turn after turn, it can beat all but the very best human players.