Measuring strength/reputation in Multi-agent systems
- Speaker: Dr Alessandro Provetti, Birkbeck, University of London
- Date: Tuesday, 18 October 2016 from 17:00 to 18:00
- Location: Room 151
Reputation is generally defined as the opinion of a group on an aspect of a person/thing.
To create and maintain reputation measures (and thus preferences/rankings) is an important stage in the development of multi-agent systems.
If we equate reputation to 'strength' then we can characterize this activity in terms of ranking players in competitive games, for which the Elo point systems for Chess is now a de-facto standard.
With Carles Sierra's group at IIIA-CSIC we have developed MORE: a reputation model based on a probabilistic modelling of opinions; it introduces a notion of information decay, and assesses the reliability of opinions based on the reputation of their sources.
MORE may exploit explicit opinions in domains where such opinions are abundant, or implicit opinions that are embedded in behavioural information in domains where explicit opinions are sparse.
We have applied MORE to sports ranking and prediction (with the development of domain-specific heuristics); the results indicate a favourable comparison with Elo predictions.