Skip to main content

Distributed Cognition Engineering

Final Report Summary - DICE (Distributed Cognition Engineering)

The project aims at the study of cognition in collective, swarm-like systems (swarm cognition), with two main goals: on the one hand, the theoretical understanding of the mechanisms that support cognitive processing and behavioural optimality in collective systems; on the other hand, the definition of an engineering methodology that provides formal methods and tools for the design of cognitive capabilities in large-scale distributed systems (e.g. distributed multi-robot systems).

These two objectives are deeply intertwined. The theoretical understanding will support the definition of a suitable engineering methodology for cognitive systems through the identification of the basic mechanisms used in distributed cognitive processing, which must be translated into formal methods and guidelines for engineering artificial systems (e.g. the interactions observed between bees during nest site selection can be distilled into design patterns leading to an optimal decision making process). Similarly, the requirements for a suitable engineering methodology can support the identification of general mechanisms used in cognitive processes (e.g. the need to provide robustness to an engineered system could point to more detailed models of distributed cognitive processing).

This challenging endeavour is instantiated in a coherent yet varied research program that starts from the study of multi-agent models of distributed cognitive processing, distills the fundamental mechanisms and proceeds to the definition of an engineering methodology for concrete applications, which will be tested on real-life swarm robotics demonstrators.

The project focuses on decision making as a collective cognitive process. Starting from models of decision making as performed by honeybee swarms when choosing the site where to build their hive, a stronger theoretical understanding of the collective decision dynamics has been achieved. The collective dynamics have been deeply studied for complex decision problems in which N possible alternative are present, and the best one must be selected. Also, adherence with psychophysical laws has been found also for collective decision performed by super organisms. We found that decision time increases with the number of available alternatives, in agreement with the Hick-Hyman law of psychophysics, and decreases with mean alternative quality, in agreement with Piéron’s law. We also found that the error rate in collective decisions depends on mean nest quality, and difference in quality, in agreement with Weber’s law. These results draw for the first time quantitative links between neural and collective decision making.

The evolutionary basis of collective decision making have been also explored. Indeed, decentralised systems such as insect colonies and animal or human groups need specific mechanisms to achieve coordination and avoid free-riding, especially when cooperation comes with a cost. Signalling is one such mechanism, but it may entail costs associated with the signal expression. Reciprocity instead requires memory of social experience, and entails an iterated process for participants to self-organise. We studied from the perspective of evolutionary game theory under what circumstances signalling and reciprocity may emerge and support collective decisions making. We treat the evolutionary process shaping self-organisation within a group in terms of an iterated n-player coordination game, where individuals can have different behaviours (strategies) and can interact through signals to reach a global group-level allocation. Our results show that there exist conditions in which collective decisions can be well supported by signalling systems, and also reveal what kind of signal should be delivered in each condition to promote collaboration and achieve coordination.

The collective decision framework has been exploited to design artificial systems, thanks to a design methodology based on design patterns. Such methodology has proven very useful to implement collective decisions in several systems, from abstract multi-agent simulations through cognitive radio networks to robotic swarms. The engineering methodology is well settled and can be exploited easily in any situation in which a decentralised decision making system is profitable.