Adapting to rules and applying new strategies is fundamental to winning any game, but this becomes much more interesting when the players are artificial intelligence agents. The folks at OpenAI decided to pit multiple agents against each other in a classic game of hide and seek. As the team of seekers began to accumulate successes, the team that had to hide incorporated the use of tools to achieve its goal, and even learned to hijack the opponent's resources.
We already know that artificial intelligences can paint pictures, clone voices, invent faces, age people and undress actresses. But the way they make mistakes and learn from them is what keeps us coming back. Of course, they have definite advantages like not eating, not sleeping, and training millions of times a day, yet their skill evolution is slow, and they have no choice but to follow that long path.
The real question is: What can they achieve when they reach the end? Cooperation. Strategic application of resources. Exploitation of established rules. The folks at OpenAI call it "Emergent Tool Use from Multi-Agent Interaction", but I think it's better if you watch them in action:
The Game Mechanics
The scenario is simple: Two teams of agents facing off in a game of hide and seek. The reward system is based on positive and negative points, +1 for "finding/escaping", or -1 if "they find nothing/are seen". With an initial streak of wins by the seekers (red), the blue team began to use and lock blocks to block entrances, creating a collaborative hideout. After several generations, the seekers also learned to use tools, moving ramps to jump over the wall and locate their opponents.
This strategy was infallible for millions of rounds, until the blue team realized that if they stole the ramps before closing the shelter, the seekers couldn't cross the wall. OpenAI didn't give the teams any kind of incentive to develop such behavior. Each time a team learns a new skill, it creates an additional pressure point for the rival, who must adapt in one way or another.
https://old.neoteo.com/google-research-football-agentes-virtuales-futbol/The Final Stage
The final phase of training moved the agents to a more open space, where the blue team would be forced to build their own shelter. The blue agents pushed the ramps off the map and locked them there, thus gaining total freedom to create an enclosed space and isolate the seekers. What did the red team of seeker agents do? They learned to "surf" on the boxes.
The Bigger Picture
OpenAI's experiment is a fabulous example of reinforcement learning. The agents trained against themselves and against previous versions using a self-play algorithm. Despite the relative simplicity of the rules in this hide-and-seek world, the agents managed to develop intelligent strategies through pure competition, collaboration, and interaction among themselves.
Official site: Click here
