Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Development of a competitive automated negotiating agent

Ziogas Ioannis-Panagiotis

Full record

---

Summary

Rational players negotiate in many real-life domains, such us the Smart Grids, health care, business and various economics environments in several. The optimal agreement to be reached from a negotiation is the one that best represents the player’s desires. A new challenge for negotiating autonomous agents is having limited or uncertain prior information about the preferences of the user they represent, as for example is demanded by the rules of the international Autonomous Negotiation Agents Competition (ANAC). This thesis aims to develop an agent which is able to negotiate given partial information about the user’s preferences, and can thus participate in ANAC. We present the design of an agent which builds mostly on the work of Srinivasan and Shocker (1973) and Tsimpoukis et al (2019). Our agent calculates its utility space from partial information based on categorical data in the form of pairwise comparisons of outcomes, instead of precise utility information. Moreover, our approach uses linear optimization for the estimation and translation of the partial information into utility estimates, by deploying the simplex algorithm. The strategy of our opponent modeling component is inspired by the Thomas-Kilmann Conflict Mode Instrument (TKI), which predicts the opponent’s strategy based on earlier counter-offers or an experience profile, as detailed in the work of Fujita (2016). Our acceptance strategy is based on the definition given by Fujita (2016), and its values corresponding to possible outcomes of the agent’s utility space. We tested our method in eleven different negotiation scenarios using four or sometimes five different domains. The strategy of our method and certain variants are deployed utilizing the well-known negotiation GENIUS platform. The negotiation tournaments that took place, involved agents with limited prior information, and also agents with complete prior information about their preference profiles. Negotiating agents that participated in past ANAC competitions comprise the agents with known preference profiles that were pitted against our agent. We conduct a systematic experimental evaluation, which demonstrates that our methodology has the ability to estimate the utility space from partial information with remarkable proximity to the real utility space. The results also verify that the algorithm’s acceptance strategy is working well while our agent accept bids based on its set acceptance threshold. Unsurprisingly, our results also demonstrate the unfairness of pitting agents with uncertain preferences against agents with known preferences. Finally, based on a close examination of our agent’s behavior during the simulations, we infer that we can improve its performance by making its acceptance strategy less cooperative.

Available Files

Services