A Method for Validating and Discovering Associations between Multi-level Emergent Behaviours in Agent-Based Simulations

Agent-based modeling and simulation have evolved as a powerful tool for modeling complex systems, offering insights into emergent behaviors and interactions among diverse agents. Recently, integrating large language models into agent-based modeling and simulation presents a promising avenue for enhancing simulation capabilities. This paper surveys the landscape of utilizing large language models in agent-based modeling and simulation, discussing their challenges and promising future directions. In this survey, since this is an interdisciplinary field, we first introduce the background of agent-based modeling and simulation and large language model-empowered agents. We then discuss the motivation for applying large language models to agent-based simulation and systematically analyze the challenges in environment perception, human alignment, action generation, and evaluation. Most importantly, we provide a comprehensive overview of the recent works of large language model-empowered agent-based modeling and simulation in multiple scenarios, which can be divided into four domains: cyber, physical, social, and hybrid, covering simulation of both real-world and virtual environments, and how these works address the above challenges. Finally, since this area is new and quickly evolving, we discuss the open problems and promising future directions. We summarize the representative papers along with their code repositories in https://github.com/tsinghua-fib-lab/LLM-Agent-Based-Modeling-and-Simulation.

Agent-Based Modeling has been recently recognized as a method for in-silico multi-scale modeling of biological cell systems. Agent-Based Models (ABMs) allow results from experimental studies of individual cell behaviors to be scaled into the macro-behavior of interacting cells in complex cell systems or tissues. Current generation ABM simulation toolkits are designed to work on serial von-Neumann architectures, which have poor scalability. The best systems can barely handle tens of thousands of agents in real-time. Considering that there are models for which mega-scale populations have significantly different emergent behaviors than smaller population sizes, it is important to have the ability to model such large scale models in real-time. In this paper we present a new framework for simulating ABMs on programmable graphics processing units (GPUs). Novel algorithms and data-structures have been developed for agent-state representation, agent motion, and replication. As a test case, we have implemented an abstracted version of the Systematic Inflammatory Response System (SIRS) ABM. Compared to the original implementation on the NetLogo system, our implementation can handle an agent population that is over three orders of magnitude larger with close to 40 updates/sec. We believe that our system is the only one of its kind that is capable of efficiently handling realistic problem sizes in biological simulations.

A number of natural and social systems often demonstrate emergent behaviours resulting from limited local interactions among their constituent parts. The emergent and self-organizing behaviours seen in many such distributed systems are closely congruent to the phenomenon of collective intelligence. Emergence and self-organization often occur together and appear to be subsumed within a broader framework of collective intelligence in distributed systems. Collective intelligence, the ability of a group to exhibit greater intelligence than its individual members, is result of various distributed interaction mechanisms, which are quite similar to the so-far identified mechanisms and underlying structures of emergence and self-organization. This paper presents a short analytical survey of the concepts and characteristic properties of emergence, self-organization and collective intelligence and tries to establish the strong similarity and inter-relatedness in them. The complex non-linear dynamics found in systems exhibiting these properties is not reducible to individual parts and therefore require a generative modeling approach of analysis. The Agent-based modeling and simulation approach seems the most suitable choice, as is also demonstrated through an example of agent-based simulation of obstacle handling in a collective setting.

In systems development, it is necessary to think thoroughly about any emergent behaviors that might occur and create strategies to prevent the undesirable emergent behavior from occurring or at least mitigate them. Often dealing with that at later development stages is costly. Emergent behaviors are the main concern, particularly in the development of complex systems. However, identifying and predicting them is challenging in complex systems as it involves lots of components and interactions. Prior to making wise design decisions, developers need to elucidate and understand the emergent behaviors. For this reason, a conceptual model that highlights the presence of emergent behavior at an early stage is essential in the systems development. The conceptual model is useful in achieving higher coverage as it is correlated with the probability of identifying more emergent behaviors. Therefore, this paper goes through the existing work focusing on effort in modeling emergent behavior and presents perspectives on developing a conceptual model that is able to inform the presence of emergent behavior.

As power markets are relatively new and still continue to evolve, there is a growing need for advanced modeling approaches that simulate the behavior of electricity markets over time and how market participants may act and react to the changing economic, financial and regulatory environments in which they operate. A new and rather promising approach is to model the electricity market as a complex adaptive system using an agent-based modeling and simulation (ABMS) approach. The purpose of an ABMS model is not necessarily to predict the outcome of a system but to reveal and understand the complex and aggregate system behaviors that emerge from the interactions of the heterogeneous individual entities. Emergent behavior is a key feature of ABMS and is not easily inferred from the simple sum of the behavior of its components. By relying on both established engineering modeling techniques as well as advanced quantitative economic market principles, the ABMS approach is uniquely suited to addressing the strategic issues of interest to different market participants as well as those of market monitors and regulators.

Software-based artificial life methods use mathematical and computational models to mimic complexity in living systems. Although such methods seem promising pertaining to exploring emergent behaviour, obtaining the governing rulesets of such methods remains challenging. In this paper, we present a concept of combined use of methods targeting different levels/scales in an emergent behaviour to obtain software-based artificial life rulesets from observed data. Additionally, we investigate the consequences of using this combination of methods by proposing an instance of combining Cellular Automata and Agent-based modelling with Koopman-based linearization. Our experiments on systems of Elementary Cellular Automaton (Rule 30), Game of Life (GOL), and Vicsek’s flocking show that the combined method can learn the overall non-linear and emergent behaviour, and the underlying governing rulesets. Our research also indicates that by identifying several emergent scales or levels in a system, the combined method has the potential to shed light on the learnt system dynamics.

Many types of financial time series, most notably market returns, have been found to exhibit long-range memory as well as dramatic day-to-day swings that cannot be adequately represented by light-tailed distributions such as the normal distribution. In particular, this means that for such time series, the usual variance parameter (i.e., the sum of covariances at all time lags) is not defined because the covariance function does not converge to zero fast enough as the time lag increases. Moreover in such time series, often the tails of the marginal density converge to zero so slowly that higher-order marginal moments such as skewness and kurtosis fail to exist. Therefore conventional methods for analyzing simulation-generated time series cannot generally be applied to high-fidelity simulations of financial markets.Building on earlier work in fractal geometry and fractal time series, Mandelbrot et al. proposed the multifractal model of asset returns (MMAR) as an alternative to the ARCH models for analyzing time series exhibiting volatility clustering, long-range dependence, and heavy-tailed returns (Mandelbrot et al. 1997). They defined the multifractal spectrum as the renormalized probability density function of the Holder exponents observed in the time series; and they used the multifractal spectrum to measure the ability of MMAR to match the statistical properties of real data. In 2002 Kantelhardt et al. formulated multifractal detrended fluctuation analysis (MF-DFA), an algorithm for extracting the multifractal spectrum from a time series (Kantelhardt et al. 2002).Many economists have recently adopted agent-based simulation for modeling financial markets. Fads based on new products, shocks related to world events, scandals involving company leaders, or outright criminal activity can drastically change the processes and relationships governing a market. Already there is evidence that agent-based models yield more accurate approximations to the true or observed behavior in financial markets. Agent-based models exhibit emergent behaviors that have been linked to non-Gaussian interaction metrics and singularities in the time series they generate (Chan 2011). Farmer et al. constructed a agent-based model that is capable of exhibiting many of the statistical properties observed in real financial data simply due to the rules governing a double auction order book (Farmer et al. 2005).During the 20th century, economists relied heavily on Brownian motion to construct theoretical models of finance. The advantages of using Brownian motion stem from its foundation in the Gaussian distribution; and with advances in stochastic calculus and differential equations, it became a staple in models of empirical finance. The well-known Black-Scholes model for options pricing is a prime example of this application. Multifractal analysis has the ability to illuminate the underlying difference between a monofractal like Brownian motion and a considerably more diverse construction like the multifractal binomial measure. Thus, combining multifractal analysis with an agent-based model has the potential to provide insight into the underlying processes and behaviors of a dynamic economic system that is constantly changing and evolving.To analyze market-return time series exhibiting volatility clustering, long-range dependence, or heavy-tailed marginals, we exploit multifractal analysis and agent-based simulation. We develop a robust, automated software tool for extracting the multifractal spectrum of a time series based on MF-DFA. Guidelines are given for setting MF-DFA's parameters in practice. The software is tested on simulated data with closed-form monofractal and multifractal spectra as well as on observed data, and the results are analyzed. We also present a prototype agent-based financial market model based on the zero-intelligence model of Farmer et al. and analyze its output using MF-DFA. The figure below compares the multifractal spectrum of the output from the prototype model with that of the multifractal binomial measure. The ultimate objective is to expand this model to study the effects of microlevel agent behaviors on the macrolevel time series output as analyzed by MF-DFA. Finally we explore the potential for validating agent-based models using MF-DFA and thus being able to tune these models to the multifractal spectrum of empirical data.

Human services planners and evaluators require an increasing high level of flexibility and adaptability to remain effective in measuring the effectiveness of social interventions. Understanding the logic and assessing the impact behind the intervention can be difficult because commonly-used evaluative tools are based primarily on linear methods that assume that a set amount of input, throughput, and output will result in a set outcome. This chapter takes a complexity science approach and facilitates the use of agent-based modelling (ABM). It provides the requisite background for evaluators and researchers to frame their efforts as complex adaptive systems. These systems have several components that include agents having options, boundaries, self-organising behaviour, different options from which to choose, feedback to adapt, and an emergent behaviour. Complexity is viewed as a mathematical field where the relations between inputs and are better understood through simulations. Both qualitative and quantitative aspects of complexity are addressed through two applications of ABM that consider related social policy issues.

In this paper, we propose a formal approach to study the emergent behaviors in multiagent systems with limited communications and environmental constraints. Temporal logic and formal grammars are employed to specify the formal model of multiagents, and based on that an automated approach can be used to identify the emergent behaviors due to interactions among agents. Instead of exploring all possible system state spaces, which may grow exponentially, the proposed approach computes the local group polarization and the local group momentum for generating the emergent behavior set, and accordingly from which the emergent behaviors can be detected. To facilitate simulation and analysis, a MATLAB GUI is further designed to enable us to flexibly select control parameters in formal models of multiagent systems. Simulation results are provided to verify the proposed approach.

Autonomous vehicles (AVs) may have a transformative influence on transportation systems. AV use is expected to reduce crash, transport costs, and land use. However, the net effects of the autonomous vehicle are unknown because there are no fully-autonomous vehicles in reality. The simulation systems are indispensable to be built up not only for cost but also for obtaining massive data. An agent-based simulation is capable of simulating the entities in a complex system, such as the AV system. Considering the potential of agent-based models to complement existing models and the difficulty to understand, to duplicate and to compare different agent-based models, this paper review existing agent-based modeling of autonomous vehicles with the help of ODD protocol, which has been applied in the review of agent-based land-use change models and agricultural policy evaluation. The significant variables that researchers have taken into consideration in sensitivity analysis and in the different scenarios are explored in the paper. Fleet size is the primary variable researchers considered, and we also discussed the probable causes of different AV replacement rates. We have assessed the quality of reviewed articles to gain more acceptance from the agent-based modeling community. Most of the reviewed papers follow the ODD protocol of agent decision making. However, in order to improve the level of modeling transparency, the executable and related data shall be available, at least the data source and the programming language/simulation platform should be provided. Several recommendations for further research are also presented.

Agent-based models (ABMs) are computational models used to simulate the actions and interactions of agents within a system. Usually, each agent has a relatively simple set of rules for how he or she responds to his or her environment and to other agents. These models are used to gain insight into the emergent behavior of complex systems with many agents, in which the emergent behavior depends upon the micro-level behavior of the individuals. ABMs are widely used in many fields, and this article reviews some of those applications. However, as relatively little work has been done on statistical inference for such models, this article also points out some of those gaps and recent strategies to address them.

Recent research discussed several approaches to understand the relation between microscopic agent behavior and macroscopic multi–agent system (MAS) behavior. A structured methodology to derive these models will have impact on MAS design, evaluation and debugging. Current results have established the description of macroscopic behavior, including cooperation, by Rate Equations derived from markovian agent–states transitions. Emergent phenomena elude these descriptions. In this paper, we argue that mesoscopic modeling is needed to provide appropriate descriptions of emergent system behavior. The mesoscopic agent states reflect the emergent behavior and allow for a deliberative implementation of the rules and conditions which cause the MAS to self–organize as wanted. In a case study, we construct such a mesoscopic model for the socio-economic inspired Minority Game. The mesoscopic description leads us to a deliberative implementation, which exhibits equivalent self–organizing behavior, confirming our results.

Increasingly, software agents are evolving from the roles of facilitators into decision-makers in managing complex, real-time systems such as logistics enterprise. Emergent behavior is that which is not attributed to any individual (agent), but is a global outcome of individual (agent) coordination. While agent's degree of autonomy and responsibility will continue to increase with time, the impact of emergent behavior in agent system on the performance and stability of these systems become an important issue. We found that emergent behavior is a major concern of potential users while we try to apply agent technology to such industry areas as transportation planning and logistics coordination. To really make agent technology work in industry convincingly, it is valuable to do a study of emergent behaviour and its impacts to the application of agent technology. This paper reviews the research on emergent behaviour, discuss its research challenge and application impact, and suggest research directions on agent emergent behaviour.

Agent-Based Models (ABMs) assist with studying emergent collective behavior of individual entities in social, biological, economic, network, and physical systems. Data provenance can support ABM by explaining individual agent behavior. However, there is no provenance support for ABMs in a distributed setting. The Multi-Agent Spatial Simulation (MASS) library provides a framework for simulating ABMs at fine granularity, where agents and spatial data are shared application resources in a distributed memory. We introduce a novel approach to capture ABM provenance in a distributed memory, called ProvMASS. We evaluate our technique with traditional data provenance queries and performance measures. Our results indicate that a configurable approach can capture provenance that explains coordination of distributed shared resources, simulation logic, and agent behavior while limiting performance overhead. We also show the ability to support practical analyses (e.g., agent tracking) and storage requirements for different capture configurations.

A deeper understanding of emergence of global patterns in social systems such as diffusion of ideas, emergence of norms & conventions, higher organizations, collective wisdom and evolution of culture; through simple and predictable local interactions of individuals, has been a long quest for sociologists. Agent-based modeling is the latest approach which has virtually replaced the use of traditional techniques of equation based models & micro simulations for social systems analysis. This new paradigm, in addition to being applied to model & analyze various social systems, is also finding widespread application in diverse domains such as economics, business organizations and computational systems. The findings of agent-based models of social systems not only help obtaining a better understanding of the investigated phenomenon but also provide valuable inputs for design of agent-based computational formulations for solving different problems in varied domains. In this paper, we have tried to characterize the use of multi-agent based modeling approach of social contagion and emergent collective behavior along with our experimental work on neighbourhood aggregation model. The paper concludes with a short discussion of the relevant implications for multi-agent systems design.