Multi-Agent-Based Simulation XXIV

Recent advances in machine learning may be leveraged by researchers in the context of agent-based modelling. With the help of surrogate models, machine learned models based on samples of a more complex agent-based model, computationally expensive evaluation methods such as sensitivity analysis and calibration may be supported and sped up. To explore the outlook on using surrogate modelling to assist simulation, possible criteria for eligibility are defined. With regards to a use case such as simulation-based crisis management and decision support, existing literature in different fields is reviewed to assess the current state of the art and potentials for holistic approaches to surrogate modelling-based simulation assistance. This work acknowledges the potentials of surrogate modelling in combination with automated machine learning, but finds no evidence that the current state of the art allows for an accessible, wide-spread usage.

Airport terminals are complex sociotechnical systems, in which humans interact with diverse technical systems. A natural way to represent them is through agent-based modeling. However, this method has two drawbacks: it entails a heavy computational burden and the emergent properties are often difficult to analyze. The purpose of our research is therefore to accurately abstract and explain the dynamics of airport terminal operations by means of computationally efficient and interpretable surrogate models, based on an existing agent-based simulation model. We propose a methodology consisting of two stages. Stage I involves the development of faithful surrogates. A sample is collected according to an active learning strategy, upon which Gaussian process regression, higher-order polynomials, gradient boosting, and random forests are fitted. Stage II then applies state-of-the-art techniques from the emerging field of explainable artificial intelligence to these models. Both model-agnostic and model-specific methods are considered, and their results are synthesized in order to explain the emergent properties. We prove the efficacy of this approach by conducting two case studies on AATOM, an existing Agent-based Airport Terminal Operations Model. Altogether, we clearly observed the preservation of emergent phenomena in surrogate models, and conclude that their combination with interpretable machine learning is an effective way to explain the dynamics of complex sociotechnical systems.

During an epidemic, it can be difficult to get an estimate of the actual number of people infected at any given time. This is due to multiple reasons, including some cases being asymptomatic and sick people not seeking healthcare for mild symptoms, among others. Large scale random sampling of the population for testing can be expensive, especially in the early stages of an epidemic, when tests are scarce. Here we show how an adaptive prevalence testing method can be developed to obtain a good estimate of the disease burden by learning to intelligently allocate a small number of tests for random testing of the population. Our approach uses a combination of an agent-based simulation and deep learning in an active sensing paradigm. We show that it is possible to get a good state estimate with relatively minimal prevalence testing, and that the trained system adapts quickly and performs well even if the disease parameters change.

We argue for a combination of declarative/constraint and imperative programming approaches for MABS: a declarative layer that specified the ontology, assumptions, types, internal and checks for a simulation and the imperative code that satisfied the statements of the declarative layer – instantiating the behaviours. Such a system would be a generalisation of common elements of existing simulations. The two layers would be separately developed and communicated but work together. Using such a system one might: (a) start by importing an ontology of entities that have been previously agreed within a field, (b) work with domain experts to implement declarative statements that reflect what is known about the system, (c) develop the implementation starting with declarative internal checks and the outlines of the implementation, (d) slowly add imperative statements to fill in details, (e) finally when the simulation has been completely verified, the declarative layer could be switched off to allow faster exploration. This would ensure for a more reliable simulation and ensure its consistency with common ontologies etc. It would facilitate: joining models together with fewer mistakes, comparing models, provide enhanced and flexible error checking, make modules more reusable, allow for rapid prototyping, support the automation of modelling tools/add-ons, and allow the selective exploration of all possible behaviours of a sub-model using constraint programming techniques. Examples are given of previous work that moves in this direction.

This paper introduces a multi-agent reinforcement learning (MARL) model for the pension ecosystem, aiming to optimise the contributor’s saving and investment strategies. The multi-agent approach enables the examination of endogenous and exogenous shocks, business cycle impacts, and policy decisions on contributor behaviour. The model generates synthetic income trajectories to develop inclusive savings strategies for a broad population. Additionally, this research innovates by developing a multi-agent model capable of adapting to various environmental changes, contrasting with traditional econometric models that assume stationary employment and market dynamics. The non-stationary nature of the model allows for a more realistic representation of economic systems, enabling a better understanding of the complex interplay between agents and their responses to evolving economic conditions (A variation of this article was included as a chapter in the PhD Thesis of Ozhamaratli, F. submitted on 22 Jan 2024).

Proper modeling of human behavior is crucial when developing agent-based models to investigate the effects of policies, such as the potential consequences of interventions during a pandemic. It is, however, unclear, how sophisticated behavior models need to be for being considered suitable to support policy making. The goal of this paper is to identify recommendations on how human behavior should be modeled in Agent-Based Social Simulation (ABSS) as well as to investigate to what extent these recommendations are actually followed by models explicitly developed for policy making. By analyzing the literature, we identify seven relevant aspects of human behavior for consideration in ABSS. Based on these aspects, we review how human behavior is modeled in ABSS of COVID-19 interventions, in order to investigate the capabilities and limitations of these models to provide policy advice. We focus on models that were published within six months of the start of the pandemic as this is when policy makers needed the support provided by ABSS the most. It was found that most models did not include the majority of the identified relevant aspects, in particular norm compliance, agent deliberation, and interventions’ affective effects on individuals. We argue that ABSS models need a higher level of descriptiveness than what is present in most of the studied early COVID-19 models to support policymaker decisions.

When social models test theories and make predictions about real scenarios, they must be fit to observed behaviors. Realistic modeling frameworks offer multiple interacting mechanisms, each with parameters that can be fit. Previously, we demonstrated how to fit the preferences that SCAMP agents use to make tactical decisions. This paper extends that work by reporting experiments on fitting the hierarchical goal networks that guide more strategic decisions.

Truly realistic models for policy making require multiple aspects of life, realistic social behaviour and the ability to simulate millions of agents. Current state of the art Agent-based models only achieve two of these requirements. Models that prioritise realistic social behaviour are not easily scalable because the complex deliberation takes into account all information available at each time step for each agent. Our framework uses context to considerably narrow down the information that has to be considered. A key property of the framework is that it can dynamically slide between fast deliberation and complex deliberation. Context is expanded based on necessity. We introduce the elements of the framework, describe the architecture and show a proof-of-concept implementation. We give first steps towards validation using this implementation.

The concept of bounded rationality has garnered substantial attention and interest from scholars since its inception. It is widely recognized that in complex systems, decision-making by its members is bounded by cognitive limitations. In this context, multi-agent simulation has emerged as a popular tool to model complex systems. One important question is how to incorporate the bounded rationality of market participants in such simulations. This paper introduces a novel multi-agent simulation model that incorporates the bounded rationality of generation companies (GENCOs) in electricity markets. We also propose evaluation metrics to quantify the differences in simulation outcomes between the proposed model and agent-based models that overlook bounded rationality, assessing the performance of market mechanisms when facing the bounded rationality of GENCOs. Using the inability of power generators to accurately predict future load curves as an illustration of bounded rationality, we conduct numerical simulation experiments on various electricity market compensation fee mechanisms. The simulation results demonstrate the effectiveness of the proposed simulation model and evaluation metrics.

Large open-source projects such as the Linux kernel provide a unique opportunity to analyze many of the socio-technical processes of open-source software development. Understanding how cognitive workload affects the quality of code and productivity of work in such environments can help better protect open-source projects from potential vulnerabilities and better utilize limited developer resources.

In this paper, we present two agent-based simulation models of developer interactions on the Linux Kernel Mailing List (LKML). We also develop several non-simulation machine learning (ML) models predicting patch reversal, to compare with our agent-based simulation models. In our experiments, simulation models perform slightly better than ML models at predicting the expected number and proportion of reverted patches, and considerably better in matching the distribution of these values. Results are further improved using an explicit process model within the simulation, modeling the patch view process and associated cognitive load on LKML reviewers when new code changes are introduced by developers. We find that the process model can capture the repeated, structured multi-agent activities within a socio-technical community.

The vehicle-to-grid feature of today’s electric vehicles suggests using them as batteries for stabilizing the power grid besides using them to fulfill mobility needs. In the context of car-sharing, the car-sharing provider may thus try to foster two goals: they may be interested in stabilizing the grid and ensuring the usage of as much green energy as possible. At the same time, they try to maximize satisfaction of the customer’s requests. As such, each car-sharing provider has to implement a policy on how to react to booking requests. On the other hand, customers may react to how mobility needs are fulfilled and adapt their booking strategy. In this paper, we study the problem of how to model elements of car-sharing providers as well as those of customers in a multi-agent simulation. We identify the principal elements and targets while leaving concrete simulations as future work.