Scalable Autonomous Network Control using Minimally Informed Agents

This thesis contributes to solving the combined problem of dynamic scalable failure risk management in complex network infrastructures while returning tractable results. It does so by proposing methods of decentralised and distributed emergent risk identification and proactive mitigation, where assets, controlled by individual agents, use only limited network information to each inform live autonomous control. This is completed in networks of telecoms - studied since information packet delivery speed is observable in computational environments - and supply chain financing (SCF) - studied since value for units is controllable. Measuring risk via criticality and viability, this thesis works around the challenge of managing a system without centralised observers, improving information delivery speeds (telecoms), and privacy (supply chains). Studies are conducted by defining simulation environments, functions to measure baseline criticality, and then investigative analyses into the relationship of baselines with decentralised or distributed approximations of the baselines. This comparison is completed for a validation function that captures distributed dynamics of the given network, with respect to network measurement or control. In telecoms, two criticality studies and one study into network changepoint analysis were conducted. The first study was direct propagation of linearly increasing criticality over a generated network. When telecoms networks cannot route data packets to an asset because it is processing other packets, they are queued. Upon asset disruption, queues are lost. Central assets must be protected because queues affect more neighbours. Graph operations on the topology of nearby assets quantify this. Restricting this to a limited region per asset can reduce response times and network pressure. The accuracy of limited measures for estimating criticality in dynamic simulations is studied, modelling criticality like an infectious disease. Accuracy benchmarking is theoretically derived. Information bounded criticality measures predict criticality. A changepoint detection study is conducted on empirical UK backbone infrastructural dynamics data, comparing two multivariate time series data aggregation methods, and of this, several changepoint detection methods for this dataset. The first aggregation method is mean aggregation. The second conducts changepoint detection on all time series, clustering changepoints using kernel density estimation, returning changepoints as maximal distances between low and high likelihood kernel density estimates. This produces a decentralised method of changepoint detection. The second criticality study proposes a model of packet routing with intercommunication, where performance of information bounded criticality estimates are measured by combining centrality measure timeliness and accuracy, creating a value of information metric. True criticality is determined using an empirical rather than theoretical benchmark. Findings show limited measures compute criticality estimates accurately and quickly, with high value of information, particularly in periodic intercommunication for degree centrality. A model of supply networks is also introduced, similar to these, using orders instead of packets. These have cash values, constraining dynamics such that policies controlling financing dynamics are implicit of criticality, replacing measurement bias with performance targets. Two studies are completed, first defining the model, then investigating the effect of information limited distributed control over a modification of it. The first defines a model of deep-tier complex SCF. A market node stimulates demand, propagating orders upstream, satisfied by deliveries in exchange for cash. Power asymmetries and payment delays induce financial stress on central but weak companies, causing failures which stress neighbouring nodes – the ripple effect. To survive this, individual agents finance themselves, dynamically changing financing benchmarks. Survival outcomes, depending on supply network topology, are explored, characterising the model. System survival times are found as bimodal, forming the basis for further investigations. Limited range centrality measures from the telecoms studies are onto an advancement of the SCF model, informing financing benchmarks. Companies only share benchmarks, thus retaining privacy and giving dependence on distant shocks with appropriate damping. Nodes thus communicate instantly while avoiding full supply chain visibility. This mitigates against ripple effects, creating competitive, indefinitely viable supply chains. The relationship between visibility and long-term survival is found, expressed as the output of polynomial regressions. Findings show that in scale free topologies, information bounded distributed control enables long-term viability. Limiting visibility information is found to outperform full system visibility for long-term survival without affecting expected survival time of short-lived systems. Thus, it is shown that, by capturing only relevant information, limited information distributed decentralised control is a scalable and effective solution to autonomously mitigate against risk in infrastructural networks.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Proselkov, Yaniv

Advisors dc:contributor.advisor

• Brintrup, Alexandra
• Parlikad, Ajith

• Complex Adaptive Systems
• Supply Chain Financing
• Collaborative Intelligence
• Visibility
• Viability
• Distributed Control
• Scalability
• Autonomy
• Telecommunications Infrastructure
• Changepoint Detection
• Value of Information
• Computational Science
• Discrete-event Simulation
• Dynamic Systems
• Intelligent Diagnostic Methodologies
• Large Scale Multi-agent Systems
• Multi-agent Simulation
• Criticality
• Complex Networks
• Ripple Effect
• Bullwhip Effect
• Financial Ripple Effect
• Bargaining Power
• Minimally Informed Agents

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/09fd7862-4735-4199-817f-51ac8baf5fbc/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0000-0002-0591-1095

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/379773