Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Exploiting linguistic data for modeling players’ behaviour in strategic board games

Apostolidou Maria

Full record

---

Summary

Many multi-agent strategic games entail social aspects realized often via natural language exchanges. Unfortunately few attempts have been made to take into account both actions and linguistic information for modeling agents. In this thesis the goal is to leverage both types of information in order to create a model that is capable of emulating players’ actions taking into account actions performed by all players in the past as well as their previous linguistic exchanges. Recent advances in neural network architectures and more precisely recurrent models allow one to sequentially update representations of the game state or linguistic data, as well as the sharing of parameters between disparate representations. Thus, in this thesis we produced and employed combined representations for the game state and for the linguistic exchanges, in order to model players’ actions and enable the prediction of their moves. We demonstrate our approach in the "Settlers of Catan" multi-agent strategic game domain. As a first step the raw data was processed to form a Dataset suitable for use in machine learning projects. This step entailed a novel modeling of the way in which information about a game of "Settlers of Catan" is represented. Then linguistic and gameplay information from the created Dataset was exploited by neural networks to predict the players’ actions. Architectures of Feed Forward Neural Networks, Recurrent Neural Networks (such as Long Short-term Memory Networks) as well as combined architectures of the two were investigated in the context of this thesis. We note that data collected in the context of the ERC Advanced Grant project STAC was used for this work, as well as the GloVe vectors for word representation.

Available Files

Services