When trying to apply recent advances in AI to a domain you often get stumped by the ‘not enough data!’ problem.
This is exactly the issue we face at Saberr when it comes to building a true AI coach for working teams. In an ideal world we would have a large dataset where the input data might contain:
Having the above gives you that unicorn called “a labeled domain specific data set”. It would be an oversized unicorn. To cover the many permutations of teams, actions, and outcomes the dataset would have to be huge.
But if you did have it you could plug it into a model, train it and tweak it as much as you want. And from that you would get a true AI team coach: one that knows the exact action to suggest for any team and context.
And getting to that point will likely take a very long time. The problem is not only in the number of data permutations but also that it needs to be longitudinal. You would need to see same data points at regular intervals over several years. The lack of trust in a team on day one may only manifest itself in a true disaster two years on.
One of the tools at your disposal when you have no data is reinforcement learning. This is where an AI agent tries a random action and then reassesses the world again to see if it got closer to its goal.
An example of this is DeepMind’s AlphaGo Zero. It would play the game of Go (against itself) many times and learn how to play it better than any human could. This happens without having labelled data in the form of human games. All that you need to know are the rules of Go to see if you are winning or losing at each step.
The trouble with team science (and social sciences in general) is that we don’t operate in a space with universal rules of the game. There is no fixed set of actions you can take and it is not clear if you are winning or losing. It would also be hard to find teams that would agree to being coached in a trial-and-error fashion.
An interesting area of research is agent-based social simulations. In a nutshell, each individual in the group is programmed to behave according to some simple rules. The groups also have group level rules that guide interactions with other groups.
This would allow us to run simulations to see what happens if a team with a given context performed a given action. For example, we could model how the addition of a new team member affects the team over time.
Combining agent based social simulations with reinforcement learning could be the way forward. The AI model would have to generate many permutations of team setups and actions during the learning phase. It would then use the simulation to see if its suggestion improved the team.
The obvious issue here is that such an AI would perform well in the simulation but may fail in the real world. Further, agent based simulations often simulate only one aspect of group behaviour, such as cooperation.
AlphaGo Zero showed the human players never before seen ways of playing Go. Perhaps this version of the AI team coach could also show human coaches radically different ways to coach teams.
Are you involved in researching team based simulations or team science in general? Would you like to discuss how we could achieve the above? Do you have better ideas? Talk to us!