Facing the challenge of automated negotiation with humans

• Autores: Angela Fabregues
• Directores de la Tesis: Carles Sierra (dir. tes.) , Josep Puyol Gruart (dir. tes.)
• Lectura: En la Universitat Autònoma de Barcelona ( España ) en 2012
• Idioma: inglés
• Tribunal Calificador de la Tesis: Juan A. Rodríguez Aguilar (presid.) , Catholijn Jonker (secret.) , Mark d'Inverno (voc.)
• Enlaces
  • Tesis en acceso abierto en: TESEO
• Resumen

  • The research field of negotiation has been studied from many different perspectives, among them: game theory, psychology, business, neuroeconomics, and psychopharmacology. The computational study of negotiations is denoted by automated negotiation. Most works on automated negotiation assume rational agents and static negotiation problems. However, humans are rationally bounded, and their negotiations are usually dynamic. It is often impossible to explore the complete negotiation space due to time limitations and the dynamics of the problem. By the time that an optimal solution is found, the solution is not optimal any more. Currently available testbeds on automated negotiation share the same shortcomings. Those testbeds that intend to involve humans in experiments assume that humans are rational, or are defined over artificial domains that require intense instruction of experiment participants. This thesis contributes to automated negotiation defining an agent architecture suitable to negotiate with humans, and a testbed that allows for an easy participation of humans in experiments.

    We denote the agent architecture by HANA. It allows multiple bilateral negotiations about actions, and deals with pre-negotiation looking for good enough sets of actions and offers. It is a modular architecture based on an ecological model of rationality. The mental state of the agent is represented as graded beliefs, dynamic desires and general intentions. We use a novel search\&negotiation technique where search and negotiation go hand in hand: the former providing offers to propose, and the latter providing commitments for pruning the search space, and information for fine-tuning the evaluation of offers. Several negotiation strategies are provided that can be dynamically combined. The architecture is extensible, allowing the incorporation of new behavioural models.

    The name of the testbed is DipGame. It is based on a popular board game where being a skilled negotiator is crucial for winning. DipGame allows the study of relationships, emotions, and coalitions that take place during successive negotiations involving humans. There are many research opportunities in different topics all of them connected to negotiation. The study of a topic or another is selected constraining the negotiation language used during the game. The testbed provides a framework for agent development, and several negotiation utilities for the representation of messages and communication among agents. It assists the execution of experiments using a graphical software application called GameManager. It facilitates the inclusion of humans with another application called ChatApp. Moreover, the analysis of results is supported by a different application called DipTools.

    This thesis is completed with a formal definition of the problem, a formal specification of the game, and the application of the work to the game industry.