Multi-level agent-based modeling - A literature survey

• The review surveys the main theoretical contributions and application domains of multi-level agent-based modeling, emphasizing social, flow, biological, and biomedical models.It addresses a research domain that has received increasing interest during the last decade.Source: Absract · p. 2

• ABM means Agent-Based Model or Agent-Based Modeling, depending on context.Source: Main acronyms used in this paper · p. 2
• DEVS means Discrete Event System Specification.Source: Main acronyms used in this paper · p. 2
• ML-ABM means Multi-Level Agent-Based Model or MultiLevel Agent-Based Modeling, depending on context.Source: Main acronyms used in this paper · p. 2

1.1 Agent-based modeling

• Classical ABM has difficulty explicitly linking microscopic and macroscopic views, spanning different scales or domains, scaling computationally, and reifying complex emergent properties.The paper notes that human observation often remains the most effective way to capture multi-level pattern formation or crowd behavior.Source: §1.1 · p. 3

1.2.1 Definitions

• ML-ABM groups approaches that extend classical ABM by adding levels of description; a level is a system viewpoint represented as a specific abstraction.The paper defines an ML-ABM as integrating heterogeneous agent-based models that represent complementary viewpoints of the same system and can interact or share entities.Source: §1.2.1 · pp. 3–4

1.2.2 Types of problems solved using multi-level agent-based approaches

• ML-ABM addresses cross-level interactions, coupling heterogeneous models, and dynamically adapting simulation detail to context.Interdependent viewpoints coexist for cross-level interaction, whereas independent models may be activated or deactivated at runtime for coupling or detail adaptation.Source: §1.2.2 · p. 4

1.3 Terminology issues

• “Multi-scale” focuses on levels’ spatial and temporal extents, while ML-ABM also concerns their interactions and organization.The paper notes that “perspective,” “interaction,” “layer,” and “view” are also used for level, and describes nested or hierarchical ML-ABMs as more precise for some models.Source: §1.3 · p. 4

2 Theoretical issues

• The surveyed theoretical work addresses meta-models and simulation engines, detection and reification of emergent phenomena, and generic representations for aggregated entities.Source: §2 · p. 6

2.1 Meta-models, simulation engines and platforms

• ML-ABM platforms and meta-models range from rule-based ML-Rules and multi-view conceptual models to frameworks such as GEAMAS, GAMA, and ML-DEVS.Their designs include different coupling and level structures: GEAMAS uses micro, meso, and macro descriptions; ML-DEVS supports hierarchies, while IRM4MLS leaves cross-level interactions unconstrained.Source: §2.1 · pp. 6–8

2.2 Observation, detection and agentification of emergent phenomena

• Emergence research aims to detect and potentially reify phenomena of interest, including for adjusting simulation detail, modeling cross-level interactions, or observing multi-level behavior.Source: §2.2 · p. 8

2.2.1 Dedicated clustering methods

• Dedicated emergence methods detect model-specific structures, including ravines and ponds from waterball movement indicators and particle clusters through graph-based methods.Other surveyed work applies clustering algorithms to identify group formation.Source: §2.2.1 · p. 9

2.2.2 Generic frameworks

• Generic emergence approaches include complex-event formalisms, reification frameworks integrated into GEAMAS-NG and GAMA, and SimAnalyzer for describing group dynamics.Source: §2.2.2 · p. 9

2.3 Representation of aggregated entities

• Aggregated-entity representations are sparsely published; surveyed approaches delegate behavioral computation to less detailed group agents to reduce interaction-computation complexity.Proposals include weighted averaging, invariant-based abstraction, pheromone fields, and mesoscopic agents; one comparison examined computational efficiency and approximation error.Source: §2.3 · pp. 9–10

• Biomedical applications include cancer, inflammation, arterial adaptation, stent design, vascular tissue engineering, and bone remodeling.Source: §3 · p. 10
• Flow-modeling applications include walking, driving, and streaming agents.Source: §3 · p. 11
• Applications span biology and ecology, with studies also covering social simulation.Source: §3 · p. 11
• Military simulation is another reported application domain.Source: §3 · p. 11

• Social theories address whether social systems should be understood through agent interactions or the structures emerging from them, and argue for considering both.Source: §3.1 · p. 11
• Reflexivity describes agents’ ongoing consideration of themselves in relation to social contexts, and of contexts in relation to themselves.Source: §3.1 · p. 12
• The section contrasts purely emergentist models of agents representing emergent structures with multilevel models that dynamically reify agent–structure interactions.Source: §3.1 · p. 12
• Practical reflexivity is linked to immergence, while discursive reflexivity is linked to second-order emergence, where agents recognize emergent properties and incorporate them into action.Source: §3.1 · p. 12
• ML-ABM links concepts at different social levels; examples include models of conflict, intra-urban mobility, and city interactions.The intra-urban model represents households, groups and urban sectors, and the city at micro, meso and macro levels; clustering detects and reifies groups.Source: §3.1 · p. 12

3.2 Flow modeling

• Traffic models distinguish vehicle interactions at the micro level, groups of vehicles at the meso level, and vehicle flows at the macro level.Micro and meso models address complex road networks such as urban areas, whereas macro models support highway congestion-control strategies.Source: §3.2 · pp. 12–13

3.2.1 Micro-macro models

• Micro–macro traffic models use a virtual section as a buffer to generate micro or macro data at connections between model types.El hmam et al. proposed coupling agent-based microscopic models with macroscopic models such as LWR, ARZ, and Payne.Source: §3.2.1 · p. 13
• Surveyed hybrid flow models have connections between levels fixed in advance and unchangeable at runtime.Dynamic hybrid modeling is needed to observe congestion formation or locate a jam within a large macro section.Source: §3.2.1 · pp. 13–14

3.2.2 Micro-meso models

• Agent aggregation can reduce interaction complexity and computer-resource use, while disaggregation is possible when the corresponding structures vanish.Source: §3.2.2 · p. 14
• A pedestrian-flow framework bases (de)aggregation on both observable environment and internal agent state, computing proximity with an affinity function.Source: §3.2.2 · p. 14
• Soyez et al. separate agents into physical bodies situated at different levels and a nonsituated mind, allowing the parts to be (de)aggregated independently.The approach was applied to adapt the level of detail dynamically in a port-operations simulator.Source: §3.2.2 · p. 14

3.3 Biological and biomedical models

• Biological systems involve different levels of organization, and explanatory models should account for interactions between levels because processes at different levels co-produce living systems.Source: §3.3 · pp. 14–15

3.3.1 Cell biology

• Cell biology models represent extracellular and cell-interaction processes macroscopically and intracellular processes microscopically, with the levels continuously influencing one another.Examples include models of Drosophila morphogenesis and Schizosaccharomyces pombe cell division and mating-type switching.Source: §3.3.1 · p. 15

3.3.2 Cancer modeling

• Cancer modeling spans molecular to population perspectives, and ML-ABM is presented as promising for exploring combination therapies across pathways.Source: §3.3.2 · p. 15
• The GBM model combines ODEs at the intracellular level, cellular discrete rules, and tissue-level PDEs while explicitly defining relations between scales.Source: §3.3.2 · p. 15
• The GBM model assigns higher resolution to heterogeneous clusters of migrating and proliferating cells and lower resolution to homogeneous clusters of dead cells, and was implemented on GPUs.Source: §3.3.2 · pp. 15–16
• Other cancer approaches include a four-level brain-tumor architecture and an avascular-tumor model that reifies the largely inactive core as one upper-level agent.The avascular model studies PAI-1 effects on metastasis and has cell and tumor-core levels.Source: §3.3.2 · p. 16

Ecology

• Different spatio-temporal dynamics complicate ecological interaction analysis; ABMs can be computationally demanding, while EBMs cannot handle complex heterogeneous environments.Ecological systems are commonly described as hierarchies, and the survey presents ML-ABM as a way to implement this perspective.Source: §3.4 · pp. 16–17
• The scale transfer approach links microscopic ABM states to an EBM of population dynamics, which can then parameterize the ABM environment.The approach is illustrated in Figure 8.Source: §3.4 · p. 17, Figure 8 · p. 17
• APSF represents soil as cells that can recursively contain pores, solids, or fractal cells, generating finer environmental detail at run time when needed.Used in SWORM to study relations between earthworm activity and soil structure, the self-generating representation is described as having minimal computational cost.Source: §3.4 · p. 17
• A Diptera development model integrates an ABM of larval crowding, a mesoscopic equation model of group temperature, and a cellular automaton for cadaver thermal dynamics.The equation model computes interaction-related temperature increases from the ABM state and passes them to the cellular automaton environment, synchronizing the models.Source: §3.4 · p. 17, Figure 9 · p. 18
• Fast multipole estimation is cited as a way to reduce the complexity of agent interactions in computationally demanding ecological models.The cited approach is proposed as an interesting solution to the interaction-complexity problem.Source: §3.4 · p. 17

4.1 Level integration

• Strong integration shares both objects and agents across levels, whereas weak integration shares objects but not agents.Weak integration resembles model coupling through shared variables or artifacts; in the surveyed meta-models, IRM4MLS and PADAWAN are identified as able to simulate strongly integrated levels.Source: §4.1 · p. 18
• Formal interaction models that distinguish agent influences from level reactions can be extended to ML-ABM because levels can themselves be strongly interacting entities.The discussion warns that modeling strong interaction as weak can lead to arbitrary implementation choices and interpretation issues.Source: §4.1 · pp. 18–19

4.2 Multi-level technical tools

• ML-ABM can serve technical purposes including ABM-to-equation-model parameterization, run-time creation of fractal environments, and mesoscopic control of group properties.The discussion also includes automated analysis levels for emergent phenomena and a multi-future approach in which ghosts compute possible action trajectories as a pheromone field.Source: §4.2 · p. 19
• Reified emergent phenomena are not model entities unless they are reinjected into the simulation.Figure 10 distinguishes visualization from reinjection as two uses of detected emergent phenomena.Source: §4.2 · p. 19, Figure 10 · p. 19
• Because these technical ideas are domain-independent, the authors suggest implementing them as a generic library for classical ABM or ML-ABM platforms.Source: §4.2 · p. 19

Conclusion

• The authors conducted a comprehensive analysis of the available ML-ABM bibliography.Source: §5 · p. 20
• Many ML-ABM papers focus on applications and appear in domain-related journals and conferences.Source: §5 · p. 20
• A dedicated venue for theoretical or methodological ML-ABM work is lacking, and the authors say one could unify vocabulary and concepts, discuss issues, and share ideas.Source: §5 · p. 20