MAS4CI - Multi-Agent Systems for Enhancing Collective Intelligence

The project proposes multi-agent systems (MASs) as a mediating technology for enhancing collective decision-making in human groups for problems admitting an objective answer. When making decisions in groups, people are affected by social biases that jeopardise the benefits of collective intelligence. For instance, people may focus on knowledge widely shared within the group, overlooking minority opinions that may be crucial. Overconfident individuals may exert excessive influence on the group, despite being no more informed than others. Herding effects may lead to groupthink without a sufficient exploration of the solution space. Similar situations are observed especially when information is openly shared within a group, making decisions excessively dependent on the contingencies of human interactions. A software application based on MAS technologies can mediate interactions among decision-makers, removing biases and channelling the decision-making process towards the right outcome.

The experiment focuses on discrete-choice problems where the best alternative must be chosen among several options. We consider problems entailing costly sampling—leading to noisy quality estimation—across multiple rounds. A classic scenario is the multi-armed bandit (MAB) problem. More complex tasks involve a perceptual decision (PD) in which an image is presented for a short time and some noisy features must be estimated. Sampling introduces an exploration-exploitation dilemma at the individual and collective level. Exploration allows to evaluate options’ quality. Exploitation allows to select (and promote) one option basing on available knowledge. Individually, an explore/exploit trade-off must be found when exploration is costly. Collectively, individuals may free-ride exploration costs taking advantage of others information. Under severe cognitive pressures (e.g., too many alternatives, time pressure), individual/social biases are expected to strike.