An algorithm created from old artificial intelligence methods triumphed against other technically more advanced programs last weekend at a StarCraft competition in Edmonton.
Despite recent advances in artificial intelligence, the “old” SAIDA has demonstrated its relevance by winning 95.91% of its 2590 matches of StarCraft: Brood War .
This 1998 real-time strategy video game features players controlling a small army on a large playing field. Players need to grow their strength by exploiting the resources on the map while trying to find and eliminate their opponents.
StarCraft is considered a game of great complexity because of the many different approaches that can be taken during the game to achieve victory. At all times, players must choose from hundreds of different options to adapt to the course of the game and the actions of their opponent.
This kind of playground is a huge challenge for artificial intelligence programs, since there is no dominant strategy and so they need to be flexible and fast enough to respond to changes in the game.
An old-fashioned method
The number of variables is so great that it may still be better to stick with the old methods that have worked in the past. That’s about what the team of South Korean Changhyeon Bae, affiliated with Samsung, has chosen.
Its algorithm named SAIDA is programmed to always use the same strategy: focus first on the defense while waiting for the best moment to hit the enemy. When a propitious moment comes, SAIDA attacks his opponent by concentrating all his forces to destroy it in a single salvo.
The program that finished in second place is derived from Facebook, while the one that finished third comes from a Chinese research group.
The technique of the future
Changhyeon Bae’s winning team is currently developing a program using Reinforcement Learning, a much more modern method that has been proven in other video games , including Dota 2 . This technique consists in giving an objective to an algorithm and letting it try thousands of strategies by itself until it reaches its goal almost every time.
Reinforcement learning has given rise to innovative tactics in competitive areas such as video games and board games. In the game of go, professional players are now training by studying the parts of AlphaZero, a DeepMind algorithm (Alphabet / Google).
The algorithms entered in the competition are still far from competing with StarCraft’s professional players , Dave Churchill, one of the organizers of the annual event , told MIT Technology Review .
The competition was held at the AIIDE Congress , which brings together specialists in artificial intelligence and interactive digital entertainment every year. It was held this year at the University of Alberta in Edmonton.