2 minute read

Last week I attended the first multidisciplinary conference on Reinforcement Learning and Decision Making (RLDM2013) and I think I speak for most people who attended when I say I hope it becomes a regular event on the academic calendar.

I often struggle to explain to people what ‘field’ I am in. To the general public I’m in computer science; to computer scientists I am in Artificial Intelligence; to AI people I focus on optimization in very large spatial MDPs using direct policy search. But in terms of philosophy and a coherent subfield I think there are a few main communities I connect with : Inference Methods in Probabilistic Graphical Models (anyone? ok…never mind), Computational Sustainability and Reinforcement Learning.

Put simply, Reinforcement Learning is the study of learning from experience, where the agent doing the learning doesn’t know how the world works or even what range of success it can expect. The agent can only rely on its experience as it goes and the feedback it receives from the world. The idea originated in psychology with Pavlovian conditional to train animals. This idea was picked up a few decades ago in computer science as a way of designing general learning algorithms and it has been very successful theoretically and more recently, in practice. But the amazing thing about this conference was that RL formed the conceptual glue which held together people from neuroscience, psychology and computer science. The psychologists are trying to build models to predict and explain human and animal learning behaviour. The neuroscientists are trying to validate biologically plausible mathematical models in order to predict the behaviour or neurons in the brain during learning. Us computer scientists are trying to build usable and effective learning algorithms for general problem solving. The magic of this conference was that we all have some common conceptual background in RL theories which we use in our research even though half of what the people from the other fields are saying is completely new and unknown to us.

But the best thing about this conference was the palpable excitement and interest in everything being presented. I heard multiple people say this was their new favourite conference.

All the computer scientists I spoke to said they learned a lot from the psychology and neuroscience talks, hopefully others felt the same way about the CS talks. There was some indication that was true especially related to the great talk by Leslie Kaelbling on learning general logical models of the world around us.

Another thing about this conference was that it had the unique property that ALL the talks were great. All of them. The organizers invited all the speakers to give half hour talks on higher level topics summarizing their sub fields and presenting context for the very wide range of people present. It worked fantastically well. I hope the organizers consider using the same model next time in at the University of Alberta in 2015. This allows the submitted work to be discussed and debated over food and drinks in a free environment. It’s nice to feel that everyone is going to all the talks because there is always something to learn from them