Periodic Reporting for period 4 - DiODe (Distributed Algorithms for Optimal Decision-Making) Reporting period: 2020-02-01 to 2020-11-30 Summary of the context and overall objectives of the project Engineers and life scientists face common problems in understanding how a collective of agents working with only local information can implement a robust and optimal collective decision-making mechanism. The DiODe project aims to contribute to the principled understanding of distributed optimal decision-making algorithms through a combination of theoretical model building, interaction with life scientists, and implementation of algorithms on a highly-scalable collective robotics platform.The detailed objectives of this project are fivefold: 1. To study optimal speed-value trade-offs in models of distributed consensus decision-making, developing novel decision theory to guide the engineering and optimisation of artificial distributed decision-making systems. 2. To study theoretically the compromise between sampling and decision-making, to predict specific mechanisms and behavioural patterns associated with solutions to such problems, and to extend the value-sensitive mechanism developed in objective 1 accordingly. 3. To develop the theory of individual ‘confidence’ in privately-held information as part of the theory of consensus decision-making, and to design distributed mechanisms able to instantiate this theory. 4. To apply the theory developed in the other objectives to the specification of individual behavioural rules for groups of hundreds of micro-robots, both to test and inform the theory, and as a new technology for robust, efficient, completely decentralised decision-making in collective robotics. 5. To test or instantiate, with named collaborators, the developed theory in natural systems including intracellular regulatory networks, neural circuits, and human groups, and to develop tools to facilitate development of consensus decision-making models by life scientists. Such tools will also be of use to engineers working with decentralised systems. Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far Under objective 1, we published further papers on optimal decentralised algorithms in leading venues including eLife, Neural Computation, Computational Brain & Behavior, and Journal of Mathematical Psychology; this work covers a variety of topics, from optimal quorum usage, to neural action selection models, to model mimicry in data fitting. The PI also published a preprint commentary on optimal value-based decisions (bioRxiv) which will be supplemented with novel experimental data for publication during the next reporting period. In the 36 month report we described how our work on generalising speed-value trade-offs to choices over more than two alternatives, including symmetry breaking and best-of-N decision-making, (described in the 18 month report) was accepted for publication in Physical Review E. It was an Editor’s Suggestion in that journal, and a research highlight in Nature Physics. We have also published a paper on replicating psychophysical, value-sensitive behaviour using the same collective decision-making model, in Scientific Reports, as described in the 30 month report.Under objective 2, work on combining sampling and decision-making in a unified formulation has continued, based on an approach pioneered by the PI in psychophysics (Cassey et al., 2013, PLoS one). Additionally, through an Erasmus traineeship to Pascal van Beek (Eindhoven) work has been conducted on analysing the optimal ongoing decision-making strategy for minimising expected cost from a decision, where cost is defined in a very general Euclidean way over a state-space following earlier work by the PI (Marshall et al., 2015, Current Zoology).Under objective 3, the review article on confidence and collective decision-making, mentioned in the 18 month report, was published in Trends in Ecology and Evolution, where it featured on the journal cover, as described in the 30 month report. While the ERC Proof of Concept grant on this was unsuccessful, the work has been further developed for publication with an external collaborator.Under objective 4, work on instantiating our theory on the Kilobot platform has progressed, with publications in prestigious venues such as ICRA, and in Swarm Intelligence; Prior value-sensitive decision-making work in Kilobot swarms was presented at Distributed and Autonomous Robotics Systems 2016 (see Publications), as described in the 30 month report. This work has built on our sophisticated Augmented Reality for Kilobots system (see §1.2) described in the 18 month report, which was accepted for publication in Robotics and Automation Letters and presented at IROS 2017, and its subsequent integration with the ARGoS robotics simulator, which was accepted for publication and presentation at the Eleventh International Conference on Swarm Intelligence (Pinciroli et al., ANTS 2018), alongside an MSc student project hosted by the project team, on value-sensitive pheromone-based collective foraging in robot swarms (Font Llenas et al., ANTS 2018; awarded as the best paper of the conference).Under objective 5, our work on our novel tool for modelling of collective behaviour systems by life scientists and engineers (see §1.2§) was released as open source (mumot.readthedocs.io) and published in PLoS one; as described in the 30 month report, this is now stored on a public GitHub repository providing a publicly-accessible executable implementation via MyBinder. In addition, project PhD student Aldo Segura has continued modelling unicellular decision-making, and has interacted with named external experimental collaborator Prof Peter Swain (see ‘Progress beyond the state of the art’). Work on theoretical neuroscience, and in collaboration with external project partners Max Wolf and colleagues, was published as described in the report on objective 1. Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far) Understanding how collectives robustly make optimal decisions has significant potential scientific and societal benefits. Improving our understanding of distributed decision-making algorithms can inform the study of decision-making motifs in diverse organisms, from intracellular decision circuits to the human brain; such understanding will be at the level of basic research, but given the importance of microbial behaviour for human health, and decision dynamics for human brain pathology, for example, the eventual societal benefit from new knowledge is potentially very large. The proposed work on optimising human committee structure could also have substantial societal benefits, for example through improving organisation of medical diagnostic committees.Impact is also beginning to be made in making sophisticated analytic tools available to non-specialists and educators, through making the Multiscale Modelling Tool Github repository public; several researchers and educators have expressed interest in using this tool.The most direct impact, however, is expected to be in the development of a new class of optimal and robust decentralised decision-making algorithms, deployed on a collective robotics testbed. To help realise this impact, a Proof of Concept grant ExODUS (Extreme and Offworld Decentralised Unanimous Swarms) was submitted during the course of the project , to attempt to commercialise DiODe research, in collaboration with robotics company GMV; however this was not funded. DiODe Logo
