When multiple AI robots and computers work at a location, they will have to collaborate with each other. They will have to work with each other just like humans do. Even if they are not working at a single location, sometimes they will have to collaborate, for example, while managing traffic, solving crimes or handling civil administration.
Recently, DeepMind, a branch of Alphabet Inc. (parent company of Google) demonstrated how AI computers could sometimes work together and sometimes couldn’t work together. In this DeepMind blog post the author explains how the AI robots are made to cooperate in different scenarios and under what scenarios they choose to cooperate or not cooperate.
One example is Gathering of Apple’s. Two robots are made to gather apples. Both the reports have the ability to tag each other – it means one robot can temporarily freeze the other by directing a beam towards it, in order to be able to gather more apples. While one robot is in the suspended state (because it has been tagged), the other robot gets more time to gather the apples. Collecting an apple triggers a reward so there is an incentive for collecting apples.
It was found that when fewer apples were left, the AI robots had a greater tendency to tag each other, which means, they decided not to cooperate with each other. Although tagging in itself does not trigger a reward but robots are able to make out that tagging the other robot somehow increases the chance of getting more apples and getting more apples is rewarding. Hence, they decide not to cooperate but hinder each other.
In another game called Wolfpack, however, how well the AI robots perform in the game depends on how well they coordinate with each other, that is, help each other. So, depending on the situation the robots may decide to cooperate more or less.
When we are cooperating in social scenarios, how we react does not simply depend on reward and punishment. It also depends on the evaluations of many bits of information and emotional cues. Will these robots be able to take decisions based on thousands of factors that affect the decisions of humans? Personally I feel it may take hundreds of years of revolution to reach at that stage.