Probabilistic Graphical Models and Serious Games

Deep-reinforcement-learning methods have lately become some of the most popular algorithms in AI, but for numerous reasons they have so far not found serious application outside of a gaming environment. In our graphical model research, we are exploring ways to use networks to bring them out into practical usage; for example, to play the “healthcare game” to find the best treatment for a patient with a complicated history or to work with practically any real-world data.

The way we handle observational data is a bridge from game worlds, where we know the rules, to the real world, where we have to tease out the rules through science and epidemiological techniques.

In our graphical model research, we seek to translate real-world processes into Markov decision processes (MDPs), which represent the change in real-world states caused by different treatments. Once expressed in this form, they can be optimized by reinforcement-learning AI and other techniques. However, in order to express data in this form, attention should be paid to how to tease causal relationships out of observational data. To do this we combine epidemiological concepts (such as the “do” function of Pearl, instrumental variables, and the potential outcomes framework) with recent developments in the new accuracy of deep neural networks.

We are currently applying these methods to a curated dataset regarding the treatments of political campaigns, and intend to next use them to address healthcare data, including data from health insurance claims. However, while these are our current foci of experimentation, the scope of potential applications is extremely broad.