New Survival Sim Demonstrates AI Can be Hostile or Cooperative

It’s a very divided topic – artificial intelligence. While many people are in full support of AI systems being integrated into society and embrace them with open arms, there are also several that fall at the opposite end of the scale. These people feel very strongly against the incoming wave of AI and see it as a real threat to the human race.

To show how these two different views could play out later on, Google’s DeepMind created a survival sim where two neural networks played against each other in two different survival scenes. The first scenario involves one where machines start behaving aggressively when resources become scarce, and the other involves working together when cooperation is beneficial. The actual scenarios were a simple fruit-gathering sim and a wolfpack hunting game. The fruit gathering sim clearly demonstrated the neural networks learning to get aggressive and shoot each their opponents a lot at times when fruit was low. But, in times where fruit was in plentiful supply, they would coexist without any issues.

It worked slightly different with the Wolfpack simulation. This time as the AI’s are rewarded for working together so is in their best interests. But, at a closer look, researchers found that two different strategies were employed when it came to the Wolfpack simulation.  Sometimes the AI would look for each other and search together, but other times one would look for the prey and then wait for the others before attacking.

Google’s DeepMind says that these simulations are ideal for demonstrating the concept of temporal discounting. This is the notion that is a reward is too far out of reach, people will tend to dismiss it. Like in the fruit-gathering sim: shooting the other players may slow you down initially, but if the reward is a chance to collect more fruit pithless competition, then surely it’s worth it?  Moving forward, Google is hopeful that neural network learning will provide a deep insight into various social science concepts and be used in applications such as traffic control, environmental science, as well as economics.

More News To Read