Agent-Based Complex Systems Science in the Era of Global Sustainability Crisis

1. Introduction

The Anthropocene witnesses unprecedented conditions and challenges about human-environment relationships [1,2]. These conditions are created by the escalating demands placed on the global environment by the largest population with the highest level of material consumption in the history of humankind. They generate challenges that range from equitable consumption [3,4] to the consequences of consumption on the functioning of the Earth system [5]. Together, these challenges have emboldened the search for sustainability—meeting the material needs of the humankind more equitably and for future generations, while not threatening the capacity of Earth system functioning and delivering the ecosystem services [6,7,8]. This search, in turn, has given rise to sustainability science, a use-inspired science seeking to advance understanding about critical elements that promote sustainable development [9,10,11]. It constitutes “a new social contract for science” [12], akin to agricultural or medical research [10], in which the approach to problem solving remains within the explanatory structure and methods of science but maintains a normative element—the goal of sustainability [11].

Human-environment interactions reside at the core of the sustainability science, and are addressed as social-environmental systems (SESs: aka social-ecological systems or coupled human and natural systems [13,14]), which behave as complex adaptive systems [15]. Comprehensive synthesis articles [9,10,11,16,17] and online repertoires [18,19] indicate that SES maintain at least three overarching elements: actors, environment, and outcome (detail in Appendix A) [10].

Several central challenges emerge in sustainability science, pursuant to its goal of sustainable development [11], that are prevalent in the synthesis articles and online repertoires noted above. It is difficult, if not impossible, to present a full spectrum of theories, approaches, advances, findings, and potential development pathways pertaining to challenges in question. Here, we focus on several broad challenges to sustainability science in which Agent Based Complex Systems (ACS) science (ACS science hereafter), as labeled by Grimm and colleagues [20], may provide potentials to resolve, especially in light of Artificial Intelligence (AI; AI hereafter). ACS science is the systems science that studies “dynamic networks of many interacting agents” [20] with an emphasis on information about entities at a lower level(s) of the system, theories about their behavior, and the emergence of system-level properties related to particular questions (detail in next section). AI, as the process of perceiving, synthesizing, and inferring information by machines [21], may substantially empower ACS science when addressing sustainability challenges. In particular, we will highlight the usefulness of machine learning (a branch of AI), which focuses on developing, understanding, and using methods that leverage data to improve the performance on some set of tasks.

The first challenge is a need to address the high dimensionality and complexity of the SES that sustainability science examines. These SES are highly diverse in kind and in the problems applied to them. Likewise, they are complex given the dimensions of factors and relationships comprising the systems [11]. The heterogeneity of elements examined at the lower and focal levels and across time draw attention to place-based or context specific outcomes, and hence resolution strategies, owing to the SES complexities, although it is understood that some kind of general processes operate throughout the system [7,11]. Given the high dimensionality and complexity of sustainability challenges, “silo approaches” [17] alone may solve one problem while exacerbating others, or relieve the problem in one dimension or moment but worsen it in others. Hence, and second, there comes the need for integrative approaches. Several frameworks for this integration have been proposed or advanced within sustainability science, foremost cast for specific problem sets common to sustainability [22] (e.g., human-nature nexus and telecoupling). At the same time, sets of metrics capturing the dimensionalities involved have been proposed, such as ecosystem/environmental services, environmental footprints, planetary boundaries, and inclusive wealth [16,23]. In perhaps the broadest framing, Clark and Harley propose that the spatial dynamics of human-environmental interactions at the mesoscale can serve as the integrator of the heterogeneity of lower-level dynamics and the relatively persistent, macro-scale patterns and processes operating on the SES [11]. Third, choosing among alternative theories or mechanisms to explain or project human decision-making or actions is a serious challenge. Alternative theories of resource uses may yield highly divergent or similar outcomes, with none outperforming the others in terms of robustness and validity [24]. Fourth and last, sustainability research and applications must enable and evaluate processes and temporal progression. This temporal dimension, including depicting and predicting pathways of development affected by hysteresis and legacies effects (i.e., lag-times between cause and effect and past outcomes constraining future ones, respectively) as well as future tipping points and adaptations in human-environmental conditions [25], becomes a must.

These challenges undergird our argument that ACS science, especially in the light of AI, may provide numerous opportunities for sustainability science. However, the sustainability science community is relatively unfamiliar with ACS science and its ABM methodology (detail in Section 2.2). The overarching goal of this review article is to illustrate the concept of ACS, its major methodology of agent-based modeling, new opportunities arising from AI, and their unique contribution to addressing the above sustainability challenges. We envision that ACS and sustainability sciences can be integrated, with strong possibilities of leading to breakthroughs in understanding and for application of sustainability problems.

2. Contribution of ACS Science to Addressing Sustainability Challenges

2.1. Handling the high dimensionality and complexity challenges

ACS Science provides a comprehensive, complex systems framework that can address the high dimensionality and complexity challenges in sustainability science. Compared to Complex Adaptive Systems [26] or Agent Societies [27] to which ACS are similar to, ACS emphasize the pivotal role of individual actors (i.e., agents in ACS) or entities (objects) that make choices (decisions) and/or act in order to pursue a certain goal [28]. Agents exist within ACS and interact with one another (Figure 1, dashed arrows) and with the environment. Agents possess different degrees of autonomy, proactivity, and intellectual capabilities, such as memory, knowledge, reasoning, learning, social capital, and adaptation. Computationally, agents are represented as software abstractions that bundle a particular set of attributes (or traits) and methods (or actions). Algorithmically, agents follow rules ranging from very simple “if-then” (reactive decision) rules to sophisticated ones based on evaluating the future consequences of alternative decisions [29]. This representation builds on a unique ontology (Figure 1) in which real-world actors are represented as heterogeneous, individual agents that comprise ACS and generate the interactions in question [30,31]. This ontology of methodological individualism represents a shift from understanding aggregate agent features and/or relationships to the individuals and micro-level processes (including interactions) that constitute and explain such aggregate features (detail in Appendix B). Given the features in this ontology (Figure 1), ACS science offers a comprehensive, complex systems framework, which can guide sustainability scientists and practitioners from the following perspectives.

First, social-environment systems under sustainability challenges can be examined in a hierarchical structure, where agents at one level or location may affect and be affected by agents at other levels or locations. To demonstrate the applicability and usefulness of this hierarchical structure, we performed a literature survey of empirical studies in both ACS and sustainability sciences (detail in Appendix C). We found that agents affect one another across lower-, focal-, and upper-levels; for example, individual migrants (lower-level agents) affect their households (focal agents) through sending back remittances [32], wastepaper markets (upper-level agents) affect decisions of wastepaper suppliers and recyclers (focal-level agents) [33]. For more examples, see Table S1 in Appendix C (under the lower-, focal-, and upper-level agent subcategories) and Appendix E, where individual monkey agents and monkey group agents affect each other across focal- and upper-levels.

Second, the importance of tracking (to a large degree) the behavior of autonomous, heterogeneous, and decision-making agents should be appreciated in SES. For instance, tracking movement of prey animals, predator animals, and hunters makes the ACS related simulation as realistic as possible: they encounter, hunt, or predate on the heterogenous landscape at certain times. This way, the simulation gives rise to meaningful results when alternative behavioral models are plugged in the agent-based model (ABM), testing the reliability of various theories of social behavior of hunter–gatherers behind these behavioral models (Appendix F).

Third, exploration of sustainability challenges should be administered in a dynamical, progressive way. This suggests that environmental conditions at earlier times may constrain those at current time, which may in turn further constrain those at future times. In this regard, there exist a plethora of case studies regarding the impacts of historic precipitation, disasters, fires, local weather conditions, and land use on the current environment (Table S1 under historic, focal, and future environment subcategories). Similarly, adjacent or distant environments may affect and be affected by the environment at the same level through various mechanisms such as the telecoupling effect [34] (Table S1 under the Same level (adjacent/distant) environment subcategory).

Fourth, decisions or actions of agents at one time or location may influence their own and other agents’ decisions and/or actions, which may translate to system-level events and/or emerging outcomes at later times or other locations. Abundant examples exist regarding how agents affect one another through crop choice, land abandonment, social norm changes, coastal defensive buildings, trading of goods, and other interactions in SES and ACS (see Table S1 under several agent-agent interaction subcategories); more discussion is in Section 2.4.

Fifth, at the system level, attention should be paid to mutual influences between SES (or ACS) across different levels, between parallel SES (ACS), or among different times. For instance, to project future human migrations and changes in the environment, Kniveton and associates point out that the interactions between parallel ACS in the future can be assessed by the exchange of information of migration destinations within a social network, which can be viewed as interconnection between the local system of migration origin and outside systems of migration destinations [35]. More examples about system-level SES/ACS interactions are presented in Table S1 (under various ACS-ACS interaction subcategories).

Finally—as a result of all the above points—this ontology provides a framework that captures the essence of many SES processes and dynamics (e.g., adaptive decision-making and the co-evolutionary aspect of ACS or SES). It guides sustainability interests in the formulation of goals (e.g., focus on focal-level alone or at focal-, lower-, and upper-levels), data collection (e.g., collect data at one time or multiple times), and analysis and modeling (e.g., perform cross-sectional data analysis, time series analysis, or simulation).

2.2. Providing an effective platform for systems integration

The modeling advances of ACS Science point to its potential in addressing the aforementioned high dimensionality, complexity, and other problems of SES and sustainability given the following considerations:

• Agents: what agents (or actors in sustainability science; see Appendix A), attributes and/or traits, and behaviors of the agents should be included at each level of the corresponding ACS or SES?
• Environment: what attributes and processes should be included (especially those affected by and feed back to affect agents) at each level? In ACS, the environment can be broadly defined to be the context other than the agent under consideration, such as the space (land) and/or other agents can be the environment.
• Agent-agent and agent-environment interactions: what relationships (expressed as rules, influences, or actions) among agents or between agents and the environment govern system dynamics at each level? What cross-level (e.g., from upper- to focal level) relationships are needed to account for systems dynamics and complexity?
• Systems-level complexity (e.g., emergence): what emerging patterns may arise from the interactions? Such patterns, often not the sum of the system’s parts, cannot be analytically solved by examination of the system’s parts alonei. This complexity includes surprises, path dependence, nonlinearity, self-organization, contingency, emergence, multifinality, and equifinality (for definitions see Liu et al. [13] and An [37]).

Sustainability science examines human-environment relationships in which actors/agents are people or people groups and the environment is the biophysical world. It seeks to understand the interactions between the two subsystems, which, more often than not, requires attention to components and interactions within or between subsystems. It is also open to applications of various methods and models, especially those that can handle integration among the components of SES [38]. ACS science, in contrast, examines any kind of relationships, agents, and subsystem interactions (e.g., bacteria and their hosts) and has heavily leveraged the use of ABMs, although cellular automata [39], partial differential equations [40,41,42], cell-based stochastic modeling [43], and structural equation modeling [44] are not uncommon (for detail see Table S1). Regardless of the range of agents entertained, ACS science provides a platform for systems integration applicable for sustainability science topics, including integration of data, information, and knowledge gained from case studies, stylized facts, role-playing games, and laboratory experiments (e.g., the four empirical approaches for social science research by Jansen and Ostrom [45]). Significantly, agent-based modeling (ABM), as a prime ACS method and tool (e.g., credited to do “a new kind of science” [46]), provides a way to fuse the deductive-mechanistic and the inductive-empirical approaches that pervade different pathways toward understanding and envisioning ACS, earning it the moniker of a “third way of doing science” [47] (see endnote ii for more discussion).

Perhaps the most advantageous feature of ABM is its capacity to provide a platform and tool for systems integration, a major goal of sustainability science [16]. Mimicking the realistic (though tailored and simplified) structure and processes of the system under investigation (Figure 1), ABM seeks to “translate” real-world actors, environment (e.g., forestland), and constraints (e.g., land use regulations; Figure 1) into virtual agents, virtual environment (e.g., land pixels), and computerized rules (e.g., if A then B else C), offering opportunities for integrating heterogenous data, knowledge, models/methods that cross spatial, temporal, and organizational scales, disciplines, and borders (e.g., political) [49] (see the exemplar ABM in Appendix E). ABMs are powerful when modeling learning and adapting processes [31,50,51], accounting for heterogeneity, bounded rationality and incomplete knowledge/information, and nonlinearities [52,53], and exploring many complexity features such as path-dependence, abrupt changes, and critical thresholds, among others [13,37].

ABMs have been widely developed and used in ACS studies to address problems confronting social, environmental, and social-environmental systems since the 1990s [55,56]. These endeavors have generated a rich legacy of ABM methodology, such as the Overview, Design concepts, Details (ODD) protocol for model documentation [57] and the Pattern-oriented Modeling (POM) approach [58] for model validation. At the same time, ABM endeavors have enriched the literature in sustainability science in terms of modeling human behavior [24,31] (e.g., the frameworks for Belief-Desire-Intentions and physical, emotional, cognitive, and social factors [27,59]), exploring how adaptive behavior, abrupt changes, crises or disasters, and critical transitions may generate surprising patterns in the corresponding SES [13,53,60], life cycle assessment [61,62], and modeling emergent macro-level outcomes and pathways under various policies or interventions [49,53,63,64].

A milestone in the sustainability science and ABM nexus was a 2006 special issue of Ecology and Society [45] addressing various empirical methods by which ABMs were empirically tested for SES. Subsequently, ABMs applied to sustainability problems have significantly increased, although they comprise only about 1.24% of all sustainability science publications in 2021 (Figure 2). Among the 29 ACS cases in our literature survey, 22 use ABM as the major method, while among the 32 sustainability science cases, only nine use ABMs (Table S1 in Appendix C). Aside from a variety of challenges in developing and employing ABMs (e.g., sharp learning curve, high data demand, programming difficulties) [31,52,55], the relative unfamiliarity of ACS science and ABMs in the sustainability science community highlights the timeliness and importance of this article.

2.3. Handling alternative pathways or theories in sustainability

ACS science has been wrestling with “finality” related challenges, which also abound in sustainability problems. Equifinality—a macro-level pattern can be generated through different pathways from micro-level processes [65]—confronts the search for mechanistic processes. In ACS science, for instance, cooperation or betrayal in the Prisoner’s Dilemma can emerge from tit-for-tat retaliation [66], strong reciprocity [67], and group selection [68], among other strategies [69]. As a double-edged sword, equifinality may offer more explanatory pathways, but also question the validity of explanations because different theories can reproduce very similar or even the same macro-patterns. In contrast, multifinality—the same causes and/or starting conditions lead to very different outcomes—also poses challenges to our understanding for mechanistic approaches [55].

The Pattern-Oriented Modeling approach [58,70], overlapping with Approximate Bayesian Computing [71] in ACS, offers a possible means to address the “finality” challenges. It is based on the multi-criteria design, selection, and calibration of models by requiring that models can simultaneously reproduce an entire set of patterns characterizing an ACS. Often a set of broad, general patterns can more effectively reduce finality issues than trying to force a model to reproduce a single pattern, such as a time series of a single variable. Given the high synergy between ACS and sustainability sciences hitherto discussed, we posit that despite its rare application in sustainability science, POM may prove useful to uncovering many sustainability related mechanisms. We refer to the example of foraging behavior model for theory testing using ABM (Appendix F).

Given the reflexivity of human agents, the social sciences tend to approach the dynamics of the social subsystem in multiple, probabilistic ways, commonly applying both quantitative and qualitative methods. Empirical models use evidence to explore outcomes and plausible, inductively derived explanations. These “top-down” models reproduce macro-level patterns that lend themselves to explanatory interpretations. For example, empirical models can accurately reproduce flight patterns of birds, even emergent ones, in the absence of theory explaining the patterns (but offering insights about the outcome to be explored). Mechanistic or “bottom-up” models, common in the biophysical sciences and some parts of the social science (e.g., economics), rely on theory-based deductive approaches. ACS science supports both approaches because its ontology explicitly represents the behavior of agents, for which theory exists and can be tested, while also providing environmental responses to that behavior and agents’ responses to the changes in the environment (Figure 1). This mechanistic and empirical blend opens opportunities to identify and explore integrated human-environment theory[38]. ACS science has empowered computational social science, allowing researchers to explore social phenomena and test hypotheses by virtue of computer-based simulations of agents and their interactions [72], nurturing a generative social science in which the dynamics are “grown” in the assessment stages [73].

2.4. Enabling and evaluating processes and temporal progression

Revealing temporal progression in the variable of interest (e.g., amount and spatial distribution of a certain resource or wildlife habitat) is important as projected patterns, if reliable, provide insights about the system’s sustainability. For instance, the dynamic habitat maps in Figure S2 (Appendix E) may inform whether the conservation policy is effective. A “byproduct” of such temporal progression information is its usefulness for model evaluation. Many investigations evaluate models (mostly statistical models) based on their goodness of fit or the maximum likelihood. Modelers strike a balance between fitting the data (e.g., by adding more parameters or equations) and keeping the explanation as simple as possible [74], reflecting the long-time trade-off between generalizability and context [45]. Evaluation of ACS models, however, does not depend extensively on statistical performance. Rather, the ACS may provide insights into the viability of the mechanistic (e.g., cognitive, institutional, and/or social) processes accounting for ACS dynamics. In this case, the ACS informs us if the processes are justifiable or not.

ACS science assists in assessing outcomes, which represent states of agents and the environment at a certain level or temporal stage, and evaluate processes and temporal progression [16], asking whether the direction, magnitude, and significance of certain parameters are supported by existing theories. In essence, all the elements and arrows in Figure 1 and Table S1 can be check points for SES documentation, assessment, or model evaluation. As the “new kind of science”, ACS science can leverage the patterns or trajectories (“data”) generated by ABM simulations, evaluating whether and how much such “data” qualitatively and quantitively agree with empirical observations or theory. For instance, sustainability researchers may consider whether the univariate and bivariate statistics or regression coefficients based on such “data” are reasonable and supported by existing theory. Furthermore, the POM approach can escalate our confidence about our understanding of the ACS and its behaviors. Finally, the ACS ontology (Figure 1) facilitates the development of new tools, platforms, or models, a high-priority research area in sustainability research [16]. For instance, An and colleagues [54] followed this ontology and developed a model to explain space-time dynamics among monkey behavior, habitat degradation, human resource collection activities, and nature reserve management policies in a Chinese nature reserve (Appendix E).

3. Opportunities from Artificial Intelligence to better understand SES

Our last section illustrates the four major advantages of adopting ACS science to address sustainability challenges. One barrier that besets both sustainability science and ACS science is the difficulty of detecting most reasonable mechanism(s) behind the data or patterns we observe, and particularly, identifying a set of justifiable rules [31,50,55]. Artificial intelligence (AI), particularly its subfield of machine learning, can substantially empower ACS [75,76]. Rather than elaborating on AI in detail, this article only aims to show the links between AI and ACS as well as their obvious implications for sustainability problems (e.g., elements in Figure 1). For this reason, our description of AI is brief, focusing on its benefits on detecting mechanism(s) behind ACS and/or SES subject to sustainability challenges.

Through a process of “training”, machine learning can help derive ACS (or SES, the ACS equivalent in sustainability science) structures or processes that verify or rebut the underlying structures, mechanisms, forces, and/or processes behind macro-patterns in the relevant ACS. Many machine-learning methods allow for the training of complex models based on some high dimensional datasets. Such machine learning methods may range from the relatively basic linear models (e.g., standard linear regression) to more advanced models that can capture non-linear behavior (e.g., neural networks, especially deep learning). On the other hand, machine learning can be used to detect patterns in model output, which may help to evaluate the robustness of the model.

Advances in data science have yielded a wide variety of scientific methods, programming tools, and appropriate data infrastructures, facilitating analysis of new forms of data (including bigdata) in a scalable, efficient, and robust fashion. This advantage boosts AI’s power to understand human intelligence and simulate how agents perceive, act, and react to other agents and/or changes in the environment(s) around them [77]. One prominent aspect of AI features neural networks, which are comprised of nodes in different layers and their links to one another mimicking human and animal brain structures. Nodes can be understood as agents in ACS or actors in SES, while links are agent-agent or agent-environment relationships in ACS or SES [78,79], which can be referred to the actors and arrows in Figure S3 (Appendix G). Once sufficient data are provided and an appropriate model structure is chosen, the trained models, often with high predictive power, help to calibrate and/or validate ACS structure or processes better. Each agent or actor can be assigned with its own unique regression equation or neural network links [80]. Understanding and envisioning agent behavior or mechanistic processes becomes a process of optimizing the neural networks for the agents iv. Recently, machine learning has advanced dramatically, helping to uncover mechanistic processes. In a successful instance [79], a graph neural network model has been trained to derive the closed-form, symbolic expression of Newton’s law of motion based on experiment data (detail in Appendix G).

Recent advances in natural language processing and mining qualitative data (e.g., ethnography input, social media

texts, and other textual sources) have shown promise to reveal the underlying reasons or explanations for a human agent’s behavior, or their stance towards a debatable issue or policy. Owing to rapid advances and the successful application of deep neural networks in natural language processing [81] and software engineering [82], it is now possible to accurately and effectively translate English text (e.g., in social media)—through developing an interactive deep learning-based system—into a list of relevant and sequential Application Programming Interfaces, which can be used to derive ABM rules or verify ABM predictions as noted in Appendix D.

As pointed out by Clark and Harley [11], “actors’ behavior and decisions, especially with respect to choices about the future, are motivated less by accurate anticipations of the future than by collectively held narratives”. Leveraging text narratives in whatever media in ACS / sustainability models can increase their potential to inform agent behaviors and/or verify outcomes in ACS [83] or trajectories related to sustainability. In Appendix D, if some “sadness” data can be collected from related tweets, ABM’s rules or predictions can be better verified or falsified about disaster or rescue dynamics.

4. Concluding Remarks

Humanity is facing a range of unprecedented sustainability challenges. Sustainability science addresses these challenges through examinations that integrate the human and biophysical subsystems that give rise to them. It blends mechanistic and empirical modeling approaches to understand the dynamics of the social-environmental systems. ACS science affords significant opportunities in these efforts. It offers sustainability researchers a unique perspective and the related means to consider relevant agents, environment, and their relationships at hierarchical levels, various locations, or times.

The contribution of this article lies in the following three aspects. First, this article points to many ACS efforts of seeking mechanistic processes, which could substantially benefit sustainability scientists. For instance, the POM approach may help better address many “finalities” challenges in sustainability science. Second, the ABM approach could offer a powerful tool for systems integration, for use of cross-scale and cross-disciplinary data and models, for model evaluation, and for providing an ontology and structure when examining a certain SES subject to sustainability challenges. Third and last, these positives are likely to be enhanced by artificial intelligence of the digital revolution (with input from data science), providing the potential to advance understanding of the social-environment systems and posit the means to make them more sustainable.