It has been almost 20 years since Deep Blue famously defeated the world chess champion Garry Kasparov, and one years since Google developed an AI which defeated the world’s greatest Go player. Now AI development has seemingly taken a further step in defeating some of the world’s greatest players in Super Smash Bros. Melee.
As Quartz reports, an AI developed by MIT student Vlad Firoiu was able to defeat some of the best Smash players in the world, representing a new frontier in how AI can develop to beat humans in every game. While the AI did not face the very best players, one of the players who lost admitted that “I am not sure if anyone could beat it.”
While this accomplishment is impressive, AI still has improvements it can make until it can actually outthink humans in every single game out there. But AIs are learning to win games beyond methods such as memorizing far more moves than a human could. They are learning to think strategically, cooperate, and actually learn how to play. This is a fine representation of how AI are learning to compete with humans and how much farther they have to go.
New environment and developments
Super Smash Bros. Melee is a 2001 Nintendo fighting game where famous Nintendo characters such as Mario and Pikachu battle one another. Despite its age as well as two sequels in the Super Smash Bros. series, it remains highly played and was featured in the largest fighting game tournament last year. Smash players are constantly using new techniques and developments to bring it to new heights.
Because Smash is a fighting game, an AI faces different challenges compared to chess. In chess, an AI can simply brute force a player by memorizing millions of moves. But in the upcoming game, Smash, an AI cannot plan ahead as it is impossible to reliably predict what a human player will do. In order to beat a player, an AI would actually have to “learn” Smash.
This is precisely what Firoiu and his colleagues taught the AI to do. They built a neural network which programmed the coordinates of game objects, and then incentivized correct play. From that point, the neural network trained against the Melee AI and then played against itself. By playing matches against itself, the AI learned in just a few weeks how to beat elite Smash players.
In addition to Smash, an AI last month defeated the world’s best players in poker through the power of neural learning as well. Just like Smash, poker is not a game won by calculating a million moves. It requires learning about a players’ tendencies and above all the ability to bluff. And like the Smash AI, the poker AI Libratus won by playing games against itself instead of human opponents, playing trillions of hands before entering the tournament. These victories in poker and video games represent how far AI has developed.
Do not rush, but be excited
But despite these victories, the way in which the AI performed in Smash shows that there is still considerable progress to be made. The AI in the Smash game was considerably more aggressive compared to a human player, but was able to get away with that partly because an AI possesses faster reaction times than a human. Furthermore, the AI has no ability to handle projectiles which are used by many characters in the game (the aforementioned matchups had the human player and AI use a character who has no projectiles) and the creator noted that the AI will actually suicide if its opponents just crouches in the corner for a while.
So it is questionable to say whether the Smash AI really outsmarted humans or just took advantage of superior reactions. However, the fact that said AI took only a few weeks to eclipse human players who had been practicing for years remains impressive. And recent AI victories in fields such as Smash and poker reveal that as a game’s parameters change, an AI can learn different kinds of behavior. This is further accentuated by two games that shows different AIs can learn to cooperate or fight each other depending on the differing natures of the game.
In a sense, anything we program an AI to do can be construed as a game. Whether it is beating players in Smash or coming up with new analysis, an AI is bound by certain parameters, faces winning or losing conditions, and can constantly improve itself whether through the actions of its programmers or through neural learning.
Throughout human history, games were often viewed as a sign of intelligence, such as how many great conquerors loved chess. As AIs are programmed to defeat human in a growing variety of games, this shows how AI can be programmed to possess different kinds of intelligence. Neural networks and machine learning have the potential to revolutionize AI intelligence, but there is still considerable progress to be made before they can surpass us in every game.