Inferring player behavior from behavioral experiments through clustering
The large numbers of participants in recent experiments where humans play the iterated prisoners dilemma game on regular and scale-free networks make it difficult to determine the actual strategies people use in playing and how they decide to switch between actions. In addition it becomes difficult to resolve whether the actions of the neighbours have any influence on the choices of a focal player. In other words how they adapt their behaviour.
One could argue that to overcome these difficulties one can use automatic inference techniques, like clustering and probabilistic modelling, to identify whether players have similar behaviour, how much their choices depend on the choices of the neighbors and which strategy-updating provides the best match to the actions observed by for the players belonging to the same group.
In this proposal you will use clustering and probabilistic graphical modelling techniques to create behavioural profiles generated from the real experimental data and simulated data.
Prior knowledge of Machine Learning and Game Theory are useful
The starting point of this thesis is the article by Gracia-Lazaroa et al:
Gracia-Lazaroa, C., Ferrera, A., Ruiza, G., Tarancona, B. A., Cuestaa, C. J. A., Sancheza, C. A., & Moreno, Y. (2012). Heterogeneous networks do not promote cooperation when humans play a Prisoner’s Dilemma. Proceedings of the National Academy of Sciences, 109(32), 1–13
In this article people have played an integrated N-player prisoners dilemma on two different kinds of networks. The aim of the thesis is to detect what kind of strategies they actually use and to model them. For the clustering we can look at:
Li, C. T., Yuan, Y., & Wilson, R. (2008). An unsupervised conditional random fields approach for clustering gene expression time series. Bioinformatics, 24(21), 2467–2473. doi:10.1093/bioinformatics/btn375
We have prior results that this clustering approach works for the data in Gracia-Lazaroa et al. Yet these results are preliminary and require further tuning.
Your contribution will be to further analyse and improve this clustering, while at the same time look for a meaningful graphical modelling framework to describe the behaviours. The resulting models will then be used in simulations to validate that they indeed result in sequences of actions equivalent to what people do. Detailed insights into probabilistic graphical modeling can be obtained through the following book (which is available in the lab):
Daphne Koller and Nir Friedman (2009) Probabilistic graphical models: principles and techniques. MIT Press
Additional analysis will be performed in data generated in a new FWO project guided by Dr. Jelena Grujic in the AI lab.
Strongly motivated students should contact Tom Lenaerts (firstname.lastname@example.org) for more information.
Supervisor: Tom Lenaerts and Jelena Grujic