My PhD dissertation summary

Today I submitted my PhD dissertation for examination. The title of my dissertation is: ‘Measuring and Influencing Sequential Joint Agent Behaviours.’

The essential thesis of my research is that:

Algorithmically designed reward functions can influence groups of learning agents toward measurable desired sequential joint behaviours. 

The thesis is demonstrated with research explaining how to measure a particular sequential joint behaviour, turn-taking, how to identify rewards that are conducive (or prohibitive) to turn-taking by learning agents in a simulated context and how to design rewards that incentivise arbitrary sequential joint behaviours in multi-agent stochastic games.

Informally, the thesis is about activities performed together through time by a group of agents that figure out how to do things better as they go. An agent could be a person, a robot or a computer program. We mathematically explain how to get the overall outcomes we want by telling the agents what they should individually want. Because we do this mathematically, we need to measure the things we want our group of agents to do. This dissertation explains some new ideas about how we can measure how well a group of agents is taking turns, how we can guess whether or not pairs of a certain kind of robot-like computer programs will take turns, and how we can tell individual agents what they should want so that they collectively end up doing something that we want, for some situations.

My dissertation includes most of two journal papers that I published, plus other bits that I’m planning to submit as another journal.

One of the things I studied was simulated agents communicating and learning from rewards.