Acknowledgement
I extend my heartfelt gratitude to my advisors, Alexis Drogoul and Huynh Xuan Hiep, along with my co-advisor, Jean-Daniel Zucker, for their unwavering support, motivation, and invaluable knowledge throughout my research and thesis writing Their organization of training programs and coding camps introduced me to the research world, for which I am immensely thankful I also appreciate the PDI program's organizers, professors, and secretaries for facilitating my participation in the international network Special thanks to Ms Patricia Zizzo for her administrative assistance at UPMC, and to Jean-Daniel Zucker, Edith Perrier, Christophe Cambier, Nicolas Marilleau, Philippe Caillou, Vo Duc An, Patrick Taillandier, Benoit Gaudou, and Kathy Baumont for their insightful advice and support Additionally, I am grateful to my professors at UPMC (Paris), CTU (Cantho), IRD (Bondy), and USTH (Hanoi), with a special acknowledgment to Dr Lai Hien Phuong for dedicating significant time to review my manuscript.
I extend my heartfelt gratitude to my thesis committee members for their valuable insights and challenging questions I also wish to thank Vo Huynh Tram, Phan Phuong Lan, and my colleagues at Can Tho University for providing an excellent research environment over the past four years Additionally, I appreciate the stimulating discussions with my lab mates at the DREAM Team, GAMA developers, and CIT, including Truong Xuan Viet and Nguyen Nhi Gia.
I would like to express my gratitude to Vinh, Truong Minh Thai, and Truong Chi Quang for their invaluable support and suggestions throughout my research, as well as for the enjoyable moments we've shared Additionally, I extend my thanks to my friends Arnaud Grignard, Nguyen Huu Tri, Tran Nguyen Minh Thu, and Nguyen Ngoc Doanh for their friendship and assistance.
I would like to express my heartfelt gratitude to my family for their unwavering support throughout my life Special thanks to my parents, HUYNH Van Sai and HUYNH Thi Nguyet, for their constant encouragement, and to my loving wife, DUONG Kim Ngan, for her understanding and affection I also appreciate my sister, HUYNH Nguyet Thanh, for her dedicated assistance during my studies.
Abstract
Computer modeling is advancing from integral to integrated models in response to the complexities of modern software engineering Today's modeling systems, particularly for decision-making in intricate socio-environmental contexts, require flexible techniques and supportive tools It is increasingly common to integrate various sub-models within a single framework, each defined across different temporal and spatial scales and expressed in diverse formalisms This integration often stems from the use of legacy models or the need to accommodate different perspectives from various fields of expertise.
Integrated modeling approaches, such as multi-simulation and multimodeling, present significant challenges in practice, primarily due to the technical difficulties of coupling diverse computational or mathematical components Additionally, aligning the semantics of these components is crucial for meaningful integration, especially in multidisciplinary models Over the past two decades, various strategies have been proposed to tackle these issues; however, none have proven to be truly suitable for our specific context, as discussed in this manuscript.
In this thesis, we introduce a novel approach known as co-modeling, which integrates concepts from agent-based modeling, agent-oriented software engineering, and multimodel ecologies A co-model is essentially a multi-agent system composed of various models and datasets, where each model or dataset is represented by one or more agents that interact within a broader, potentially dynamic framework.
The proposed approach offers a framework that allows modelers to easily implement and test various coupling solutions, rather than providing a one-size-fits-all solution to the identified challenges in the environment.
The proposed approach is implemented in the GAMA agent-based modeling platform, showcasing its flexibility, composability, and reusability across various case studies The first case study demonstrates the dynamic coupling of equation-based and agent-based models to create "switching" models The second case study highlights the successful integration of three formalisms and four modeling approaches into a complex model.
2.4 Existing solutions in modeling/simulation 35
AN INFRASTRUCTURE FOR SUPPORTING THE DYNAMICAL
3.3 Integration in the GAMA platform 51
4.4.1 Step by step co-modeling the prey-predator co-model 65
5 Dynamic choice of the best representation of a phenomenon 72
5.4 Transformation of the Switch model into a co-model 81
6 Incremental design of a complex integrated model 85
6.3.6 Summary on the micro-models 96
6.4 Impact of environmental factors on farmers’ decisions 97
6.5 Coupling of Farmer and Socio-Economy factors 100
6.6 Coupling environmental, social and economic models 102
Multi-modelisation is not different with the modelisation It does not need a new language to describe the coupling of models.
Context
Complex systems encompass a wide range of fields, including life sciences, medicine, physics, chemistry, engineering, and social sciences Research in this area necessitates an interdisciplinary approach that transcends traditional boundaries In these systems, the interactions among individual components lead to emergent functionalities that are not observable at the individual level, posing a significant challenge for contemporary science To effectively model and simulate these complex interactions, a multi-modeling approach is essential, as different interactions may belong to various disciplines The resulting models, known as "coupled" or "integrated" models, consist of multiple sub-models, some of which are "legacy models" created for specific questions in different contexts Designing these coupled models remains a significant challenge for researchers working on complex systems today.
“coupling” can represent any link between the different sub-models: dynamic interactions and feedbacks, static or dynamic compositions or combinations of these models in various frameworks.
Different definitions of complex system could be found in the literature:
A complex system consists of numerous interacting entities or processes, necessitating the use of advanced scientific tools, nonlinear models, and computer simulations for a comprehensive understanding.
A complex system is composed of numerous interrelated components, where the behavior of each part is influenced by the actions of others This interconnectedness highlights the intricate dynamics that define the overall functioning of the system.
A system comprises multiple interacting agents, and understanding their collective behaviors is essential This aggregate activity is nonlinear, meaning it cannot be accurately predicted by merely summing the behaviors of individual components.
This thesis was conducted through a collaboration between the IRD UMMISCO research team and the University of Can Tho in Vietnam, established as a JEAI (Jeune Equipe Associée à l’IRD) known as DREAM (Decision-support Research for Environmental Applications and Models).
In 2012, I collaborated with researchers through the “Programme Doctoral International Modélisation des Systèmes Complexes” (PDI MSC) at Pierre-and-Marie-Curie University and the IRD, focusing on simulation projects that involved the design of complex coupled models One project aimed to develop evacuation policies for tsunami scenarios in Da Nang city, while the other focused on creating information systems to protect against epidemic diseases affecting crop plants and aquaculture in key economic areas of Vietnam These complex models, which integrated multiple sub-models, are detailed in the following sections.
Da Nang, a major port city in Vietnam and the largest on the South Central Coast, is vulnerable to tsunamis from the Southeast Asia Sea, as highlighted by the aftermath of the Tohoku earthquake In response, the People's Committee of Da Nang has pioneered the installation of 10 early tsunami warning stations in Vietnam This initiative is supported by the Vietnam Institute of Geophysics, the National Committee for Search and Rescue, and local authorities, who conducted a tsunami drill involving over 6,000 participants The drill simulated a scenario where a magnitude 8.8 earthquake off the Philippines triggered a 6-meter tsunami, expected to reach Da Nang's coast and advance 700 meters inland within two to three hours, with 6,000 tourists and 1,000 fishermen at sea.
The drill was successful and the directive board decided to build tsunami scenarios and plans to cope with this kind of disaster.
Several questions could be raised and used as attractive research topic:
• How do the actors understand the risks caused by tsunami?
• What are the most relevant indicators of “community resilience” for these risks?
• What are the existing policies and recommendations for building this resilience?
• What are the conditions for stakeholders to implement these recom- mendations?
• Are the local communities trained to put these recommendations into practice?
A project integrating a multi-agent paradigm with data analysis has shown promising results in tsunami evacuation planning The simulation model views tsunamis as destructive waves, while the evacuation strategy aims to maximize survival rates This plan considers various factors, including local government policies, available rescue resources, and the dynamics of urban infrastructure during disasters It incorporates geophysical models that represent elements triggering earthquakes or tsunamis, alongside a government evacuation plan that can be data-driven or computational By coupling these models, decision-makers can evaluate and optimize evacuation strategies regardless of tsunami type or intensity, ultimately creating a comprehensive model to assess hazards, vulnerabilities, risks, and evacuation solutions in tsunami scenarios.
In the Mekong Delta region of Vietnam, agricultural managers are increasingly concerned about the invasions of Brown Plant Hoppers (BPH), a rice pest that carries diseases detrimental to rice yields Accurate estimation of BPH distribution is crucial for developing effective prevention plans, as their spread is influenced by various environmental factors, including wind speed, humidity, temperature, urban development, changes in rice fields, and climate change BPH invasions present a complex spatio-temporal challenge, characterized by a short life cycle consisting of egg, nymph, and adult stages, and the ability to migrate via dominant winds in search of food sources External factors, such as farming practices, pesticide use, and weather variability, further impact BPH populations This complexity necessitates a multidisciplinary approach in modeling BPH dynamics, as demonstrated by researchers at Can Tho University, who integrated multiple social, biological, and land-use models to better understand the interplay between BPH invasions and control policies.
Motivations and Research questions
Current solutions like High Level Architecture (HLA) offer various model couplings but fail to adequately address the challenges outlined in Chapter 2 Many of these solutions share similarities with modern software engineering practices, yet most modelers in fields such as social and ecological sciences are not trained as software engineers Despite this, over the past two decades, there has been a growing adoption of computer modeling formalisms, particularly Agent-Based Modeling (ABM), which has become prevalent in exploring interactions among diverse components in complex systems While ABM provides a dynamic composition of sub-models called agents, it lacks a standardized methodology for model coupling This thesis aims to leverage the widespread use of ABM to demonstrate its advantages for model integration, allowing modelers to create integrated models without needing to master new concepts The manuscript will address significant challenges, including the complexities of coupling heterogeneous models and aligning their semantics for meaningful integration, particularly in multidisciplinary contexts.
This thesis presents a comprehensive solution known as co-modeling, designed to tackle significant challenges in the field It enhances the agent-based modeling paradigm by providing a cohesive approach to multi-modeling and model coupling The conceptual framework outlined in the following chapters has been fully implemented and tested as an extension of the GAMA agent-based modeling and simulation framework This implementation has been evaluated for its flexibility, composability, and reusability through various case studies.
The organization of this manuscript is as follows:
Chapter 2 explores the latest developments in model coupling, addressing the needs of the modeling community and identifying the limitations of existing methods It emphasizes the application of models in multidisciplinary research areas such as urbanization, traffic management, and socio-environmental systems, highlighting key requirements for modelers in these fields The chapter examines various proposed formalisms and frameworks used for model coupling, comparing their offerings against the identified requirements A comprehensive synthesis of their respective advantages and disadvantages provides insights into the current options available for constructing integrated models.
Chapter 3 introduces a novel agent-based approach to model coupling, termed co-modeling, which is implemented in the GAMA computer modeling and simulation platform This approach builds on the understanding that modelers, while not software engineers, are increasingly adept with computer modeling techniques like agent-based modeling (ABM) The chapter outlines the key concepts, syntax, and operational model of co-modeling, emphasizing its advantages in terms of inclusivity, scalability, and reusability, thereby enabling ABM-experienced modelers to create intricate coupled models effectively.
In Chapter 4, we implement co-modeling using a straightforward integrated model, the Prey Predator toy, highlighting its benefits compared to other methods, particularly regarding expressivity and flexibility.
Chapter 5 illustrates the benefits of utilizing agents for model representation, emphasizing their role in enabling modelers to create dynamically composed models that can adapt their representation to the systems being modeled.
Chapter 6 explores the potential of co-modeling by showcasing the incremental design of a complex integrated model This chapter illustrates how diverse and heterogeneous models can be effectively combined to address questions related to the management of intricate social-environmental systems.
• Chapter 7 concludes this thesis by analyzing the outcomes of our research over these past 4 years as well as the perspectives for future works.
This chapter defines coupling in modeling and reviews current approaches to model coupling It examines the challenges of integrating heterogeneous models using existing methodologies and tools, as well as the needs of the modeling community for multidisciplinary models By providing a thorough overview of existing methods, we highlight their strengths and weaknesses, leading to the proposal of an initial design for a more flexible solution.
Why coupling models ?
Model
The term "model," particularly in the context of computer modeling, has various interpretations A model is defined as a computer structure that extracts essential attributes from a system Typically, models are constructed using a subset of an original prototype’s attributes and can serve as substitutes for the prototype in specific scenarios The selection of these attributes is influenced by the questions that modelers aim to address; for instance, in ecology, modeling seeks to encapsulate the core aspects of a system to answer particular inquiries The complexity of the systems being modeled is constrained by researchers' observations and knowledge, as well as the scientific questions posed In cases where the inquiry involves the interactions of multiple sub-systems—such as social and biological systems in epidemiology or hydrological and urban models in evacuation planning—there may be a need to develop "heterogeneous models." These models integrate various target systems and consist of multiple components, each potentially built independently.
Models Coupling
As questions posed to models grow increasingly complex, model coupling has become more prevalent, particularly in sustainable development where multidisciplinary collaboration is essential This coupling is often driven by the need to integrate diverse models, such as urban and climate models or flood and evacuation models, and to select the most suitable ones for specific applications Additionally, it facilitates the understanding of phenomena across various spatial and temporal scales and allows for the incremental modification of model structures to adapt to evolving questions and requirements in decision-making processes.
In general, models are said to be “coupled” (and not simply, for instance,
Models are considered "juxtaposed" when they can function independently to address specific questions and when their integration into a larger framework involves interactions such as data exchange and control flow Coupling is essential for analyzing complex heterogeneous systems that require multiple levels of detail, often necessitating the combination of existing models This approach is particularly relevant for interdisciplinary inquiries, such as assessing the effects of climate change on socio-environmental systems, which demand collaborative research across various scientific fields and consideration of different spatial and temporal scales.
Advantages of coupling
• Reusing models allows to cut down the costs and efforts to develop new models In the disaster evacuation case study described in Example
The geophysical model provides critical data on tsunami probabilities and their potential impacts, such as wave height and effects on infrastructure This information is utilized by an evacuation model that evaluates existing government evacuation plans to identify the most effective strategy The evacuation model simulates individual behaviors and interactions as people attempt to reach shelters quickly Each of these models has been developed by experts in their fields and integrated to create a comprehensive system, avoiding the redundancy of redeveloping similar components like water levels, buildings, and population dynamics.
The spread of brown plant hoppers (BPH) in the Mekong Delta is influenced by various factors across multiple spheres, including environmental aspects like wind speed, humidity, and temperature; social elements such as control policies, early-warning systems, and farmers' practices; and biological factors involving BPH predators and alternative crop species Different domains study these spheres, leading to the development of numerous specific models at various spatial and temporal scales Due to the complexity of the interactions between the environment, human settlements, and the insects, models for assessing and predicting BPH invasions must integrate existing models effectively to address new challenges, such as optimizing surveillance networks and control policy mixes.
Coupling models from various domains enhances the realism and comprehensiveness of system representations, ultimately reducing development costs This integration allows modelers to include essential information that may be missing or outside their expertise, such as tsunami dynamics for evacuation planning An evacuation planning expert may lack the ability to analyze such complex phenomena but requires accurate data for effective model design Additionally, the unique behaviors of individuals during evacuations can significantly impact outcomes, necessitating the use of individual-based models to validate evacuation plans By improving model realism through multidisciplinary collaboration, it becomes possible to enhance the adaptability of evacuation strategies, optimize resource management, and ultimately save lives.
Coupled models offer enhanced analysis, understanding, and interpretation compared to complex all-in-one models by allowing a clearer separation of concerns and greater independence of parameters This approach improves model readability, documentation, and maintenance over time For example, in the evacuation case study discussed in Chapter 1, the geophysical model involves over twenty parameters, with minimal interdependence among them By isolating these parameters in separate models, we can validate them independently, facilitate reuse in different contexts, and better understand their influence on other models Additionally, a coupling approach promotes efficient exploration of integrated model parameters and fosters better organization of multidisciplinary collaboration by clarifying the intersections, dependencies, and interactions among models and their respective experts.
Different kinds of coupling
Weak coupling
An illustrative case of weak coupling is demonstrated in the integration of UrbanSim and MATSim to create a comprehensive urban mobility model UrbanSim offers insights into residential areas, job locations, and urban development, while MATSim contributes extensive data on land use, transportation networks, and economic dynamics The synchronization between these models occurs exclusively through data exchange, with UrbanSim sending mobility needs to MATSim and receiving accessibility indicators in return Other examples of such weak coupling can be found in existing literature.
An integrated marine environment model consists of local ecosystems of pelagic species, each functioning independently yet interconnected through shared input and output parameters For instance, a study couples a community land model with the West African monsoon's regional climate model However, simple data exchanges are often inadequate in weak coupling approaches due to the differing scales, objectives, and potential legacy constraints of sub-models To address these challenges, a "coupler" is necessary to translate the unique characteristics of each model, which may involve various formalism types such as agent-based modeling or continuous equations Numerous studies have explored the coupling of models from different paradigms, including hydrodynamic and individual-based models, as well as physical and social models, demonstrating the complexity of integrating diverse ecological frameworks.
In the case studies presented in section 1.1 of chapter 1, modelers predominantly employed a weak coupling approach This choice in the first case study was influenced by specific constraints, such as the geophysical model's construction.
In our classification of coupling models, we distinguish between two types: weak coupling and strong coupling Weak coupling refers to systems that primarily accept input parameters and yield numerical outputs, such as sea level measurements, while strong coupling involves more integrated models where components can be modified The models are represented by red-rounded rectangles in Figure 2.2.
In the second case study, the modeling choice involves utilizing the number of BPHs along with wind, temperature, and humidity values to couple various sub-models This approach ensures that the sub-models, which would otherwise operate independently, can effectively communicate, thereby enhancing their reusability in different modeling scenarios.
Strong coupling
In certain scenarios, executing or simulating sub-models requires enhanced control, particularly when integrating various models of the same phenomenon where only one model should be utilized at any given time or spatial scale This need is further emphasized when running stochastic models repeatedly to refine the confidence intervals of their outputs In such instances, a stronger coupling is essential, which incorporates functional controls alongside data exchange Strong coupling typically employs operational architectures that facilitate control over sub-models, utilizing existing modeling paradigms, such as agent-centered approaches that translate different modeling formalisms into agent-based models, or specialized software architectures designed for functional model coupling, including High-Level Architecture (HLA), Discrete Event Systems (DEVS), and Functional Mock-up Interface (FMI).
Both strong and weak coupling approaches have their pros and cons Strong coupling allows for more integrated solutions but can hinder flexibility, making it difficult to replace sub-models Conversely, while a weak coupling approach offers greater flexibility, it necessitates the design of an interface that may restrict the types of sub-models that can be utilized.
This thesis advocates for a balanced approach that integrates both weak and strong coupling dynamics, offering modelers greater flexibility and versatility In the following section, we will outline the specific challenges related to model coupling that our approach aims to tackle, as well as evaluate the successes and shortcomings of existing methods in addressing these issues.
Challenges of model coupling
Reusability
Reusability in software engineering encompasses both technical aspects, such as linking libraries across different programming languages, and conceptual elements tied to "controlled genericity," where software is designed for reuse In model coupling, the challenge of reusing existing models mirrors these issues but introduces a semantic constraint related to the formalism used in model expression Integrated models often employ multiple, conflicting formalisms, complicating the translation between their representations For example, a model may integrate differential equations, cellular automata, and agent-based approaches, each with distinct meta-models Thus, reusability in model coupling focuses more on the transparent integration and translation of sub-models within an integrated framework than on the technical feasibility of such integration.
Scalability
When integrating different models, modelers encounter scale translation issues due to varying time and space scales or specific input requirements For example, equation-based models typically function at aggregate scales, while agent-based models focus on local scales This discrepancy necessitates a clear understanding of the operational levels of each model and the flow of data and control between them A notable example is the differing spatial scales in evacuation models, such as those for tsunamis and evacuees, highlighting the complexity of model coupling.
The same model can be applied at various scales by adjusting parameters and interpretation accordingly In the BPH invasions model, static light traps are utilized to capture and count insects, with their zones of influence modeled at both the town/district level and larger scales, such as quarter, province, and region To effectively implement the model across these different contexts, spatial discretization and aggregation operations are essential.
Translating data between different temporal scales of sub-models is essential for effective integration For example, meteorological data from BPH invasion models is collected monthly, while trap density data is gathered daily This discrepancy necessitates a translation process, which cannot be entirely independent of the models involved In some cases, averaging monthly meteorological data may suffice, but in others, where specific weather events like rainfall are critical, a more nuanced approach is required.
An effective and scalable coupling mechanism must facilitate clear expressions of the translations between the spatial and temporal scales of each sub-model, regardless of their complexity This is crucial, as these translations may become integral components of the integrated model, similar to other sub-models For instance, when dealing with monthly meteorological data that includes average rainfall, it is essential to determine the appropriate timing for simulating rainfall events on a daily or weekly basis.
Expressivity
An integrated model should be viewed as a collection of sub-models and the connections between them to ensure reusability and scalability This necessitates a coupling infrastructure that employs a modeling language to describe these connections, similar to how agent-based models use specific languages for agent interactions or mathematical models utilize equations for variable relationships Ideally, this modeling language should be easily comprehensible to modelers, leveraging their existing expertise in model design and development However, many current solutions distinctly separate the languages used for expressing models from the coupling language.
Modelers often face the challenge of learning entirely new languages and concepts to describe couplings in multi-simulation frameworks like HLA and FMI These powerful tools necessitate the use of structures and terminologies that are not typically associated with traditional modeling practices, complicating the process of even simple couplings.
Flexibility
A key requirement for an effective coupling approach is its flexibility, which is essential not only during the design phase but also throughout model exploration, experimentation, and testing during simulations This flexibility allows for dynamic changes to the integrated model's structure at runtime, such as swapping sub-models, adding or removing models, and modifying coupling configurations Such adaptability is crucial in integrated models that incorporate learning mechanisms or when multiple sub-models can fulfill the same role, albeit with varying accuracies and requirements For example, historical datasets may provide necessary data for a specific timeframe, while models can be utilized to fill in gaps when this information is unavailable.
Existing solutions in modeling/simulation
Many existing integrated models utilize "ad-hoc" coupling techniques tailored for specific sub-models, making them difficult to reuse in different contexts For instance, the integration between UrbanSim and MATSim is uniquely designed for that particular case, preventing adaptation or extension to include other models, such as environmental ones This challenge often arises when models are drawn from diverse domains, such as urbanization and transport, or environmental studies and pelagic resources.
Terrestrial ecosystems are significantly impacted by regional climate change, particularly when factors originate from the same domain Data exchanges within these systems often occur in an ad-hoc manner, despite attempts at standardization in integrated models One such model employs an explicit coupling mechanism to enhance data integration and coherence.
“coupler”, that exchanges input and output data between models operating at different spatial and temporal scales.
In conclusion, various modeling approaches can be effectively coupled, such as hydrodynamic models with individual-based models, multi-agent models with GIS models, and physical models with social models However, these "ad-hoc" coupling methods lack generalizability to other modeling contexts or problems, limiting their broader application.
The challenge of complex model coupling has prompted the modeling community to develop various architectures and frameworks, such as the High-Level Architecture (HLA), Discrete Event Systems (DEVS), and Functional Mockup Interface This article will present and analyze these three robust examples to determine if they fulfill the necessary requirements for effective model integration.
High-Level Architecture (HLA), originally developed for the military sector, is primarily utilized for human training through task performance and scenario analysis in simulated environments It facilitates the synchronization of heterogeneous simulators during data exchange, with the core principle being that simulators operate as part of a Federation The Runtime Infrastructure (RTI) interface ensures seamless communication between these Federations HLA comprises three essential components: the template object model (which includes the HLA Federation Object Model and HLA Simulation Object Model), the interface specification with RTI, and HLA rules Despite its advantages in managing diverse simulators and a robust implementation, HLA's complexity poses challenges for non-computer scientists, limiting its accessibility for many modelers.
Discrete Event System Specification (DEVS) is a formalism designed for modeling discrete event systems, notable for its recursive definition that allows models to be classified as either "atomic" or "coupled." An atomic model includes essential parameters such as input and output events, sequential states, time advance, and transition functions In contrast, a coupled model incorporates additional features, including the set of atomic models it connects, a translation function for input-output relationships, and specific ports to manage interactions between models While DEVS offers an elegant framework for describing model coupling and is effective for creating composite models, its reliance on a formal and deterministic structure limits its applicability in assembling stochastic or complex legacy models.
The Functional Mockup Interface (FMI) is an industrial standard designed for co-simulation, allowing each sub-model or simulator to interact through a defined functional interface This independent model exchange approach facilitates 'black box' exchanges, meaning users do not need to understand the internal workings of a model as long as its interface is well-defined Developed to meet industry standards for standardization, accessibility, ease of use, and model coupling maturity, FMI remains largely underutilized in academia, despite its beneficial features, particularly in standardization However, its software engineering roots introduce a unique model description method that diverges from traditional modeling paradigms found in environmental and social sciences.
HLA, DEVS, and eling are operational tools used to tackle environmental and socio-environmental issues, yet their complexity poses challenges for modelers These frameworks primarily focus on software engineering aspects of model coupling, necessitating a steep learning curve for users Consequently, academic modelers may find the requirement to redesign or rewrite sub-models to fit these coupling infrastructures daunting, while industrial practitioners might be more open to such extensive modifications Additionally, the formalisms and languages employed by these techniques differ significantly from those typically used in developing socio-environmental models, such as agent-based models, Cellular Automata, and mathematical models.
In addition to the three standard solutions, the modeling community has seen various contributions from designers of simulation platforms, each presenting unique coupling methods and definitions for spatial, temporal, and data exchanges, often specific to their platforms The key advantages and disadvantages of the primary solutions are outlined in the table below.
- It is a good conceptual approach for describing the assemblage of possibly interacting atomic sub-models within a larger coupled model
- It is mostly a conceptual approach, implemented in a few tools.
- It offers a very good support for designing complex discrete-event models and simulations.
- It is difficult to describe models that do not follow discrete-event dynamics (continuous time, discrete time-based models).
- It does not really offer supports for complex behavioral or environmental declarations.
- It is very limited when it comes to couple stochastic models.
- It is easy to reuse and combine simulations in an interoperable manner.
- It is a pure technical solution in term of a standard that does not support dynamical changes in the coupling structure.
- Distributed simulations are well controlled with data communicating and synchronization of actions.
- It does not focus on the modeling of the coupling, which is disseminated in the various structures.
- It mostly requires to re-implement legacy models in theC++ language or to build quite complex wrappers around them.
- It supports the coupling of physical models emanating from different domains
- Co-simulation interfaces are only available (and documented) for the engineering and industrial domains
- It allows exporting and importing model components in industrial simulation tools (FMI for Model Exchange)
- It standardizes co-simulation interfaces in nonlinear system dynamics (FMI for
- Its support in terms of modeling is quite limited, i.e it mainly focuses on the co-simulation and exchanges of models, not really on the description of their coupling.
- It provided the most efficient co-simulation interfaces between electronic, mechanical and software models The primary goal of FMI is to simulate and analyze these models.
- It supports the definition of quite complex and dynamical organizational structures between software components called agents, similar in practice to the agents used in ABM
- It is not specifically dedicated to modeling (in spite of the presence of a modeling language called TurtleKit, similar to the one found in NetLogo)
- It uses a recursive definition similar to DEVS but less limited
- It does not offer any support in terms of multi-formalism (besides the ability to write agents in Java, Python or C++)
- It is an effective method for coupling many parallel models to form one high-performance coupled modeling system.
- It is dedicated to atmosphere and ocean general circulation models, land-surface models, and dynamical sea-ice models.
- It uses a set of outdated Fortran90 modules defining derived data types and routines
- It is an interactive modeling GUI that provides an easy access to the DEVS framework.
- It does not have a specific language dedicated to modeling.
- It does not offer a lot of support for multi-formalism models or various data sources.
- It suffers from the same problems than DEVS regarding stochastic models.
- It is the most popular modeling platform
- It allows modelers to develop only medium and small models, mostly for academic purposes.
- Its model library is plentiful and covers most of the research domains.
- It does not support the coupling of models, even simple ones.
- It is a powerful code-generating tool for multi-formalism modeling.
- It only works with models that can be transformed into graphs.
- Instead of building the whole application from scratch, it only requires to specify, in a graphical manner, the types of models that need to be coupled.
- Coupling is supported by agents which encapsulate existing softwares as components of agents.
- It is not adapted, due to performance problems, to large models.
- It is not really maintained anymore.
OSIRIS[25] - It is flexible and generic.
- The sub-models cannot be modified at runtime
- It allows to couple different physical processes.
- It does not support explicit spatial coupling.
- Its GUI helps to define the characteristics of coupled models.
- It is a high-performance software platform for delivering solutions targeted to complex business problems.
- It does not have a specific language dedicated to modeling.
- It does not offer a good support for the description of spatial processes.
Conclusions
While the methods discussed in this chapter facilitate varying levels of interaction between sub-models in integrated models and allow for some expressiveness in their coupling descriptions, they also possess certain limitations.
Many existing standards, aside from the three primary ones—HLA, DEVS, and FMI—lack the flexibility needed to define coupling architectures that can be readily reused across various domains or contexts.
• Most of them, besides HLA and DEVS, do not support the definition of explicit spatial and temporal scales and the translation between them.
When modeling is not restricted to a specific paradigm, such as DEVS, it can accommodate a wider range of models, including stochastic ones This flexibility allows for the use of familiar languages and concepts, making it easier for modelers to engage with their work effectively.
• Finally, the majority of them (besides DEVS and possibly HLA) have never been tested against the design of complex data-driven, multi-formalism integrated models.
This article explores innovative concepts in model coupling, highlighting structures like ARCHON and OSIRIS for encapsulating legacy models, DEVS and MADKIT for recursive organization, and the flexible ‘temporal controller’ from HLA These designs address the challenges of model integration and should be adapted in future proposals Building on these existing works, I introduce a new approach to model coupling in the following chapter, which aligns with the "multi-model ecologies" concept and features a comprehensive agent-based implementation in the GAMA platform This proposal not only showcases various integrated models but also aims for broad applicability, representing the key outcome of my research.
This chapter addresses the limitations of existing models coupling methods and proposes a novel agent-based approach called co-modeling, implemented within the GAMA platform Recognizing that modelers are not necessarily software engineers but are increasingly adept with computer modeling techniques like agent-based modeling (ABM), the chapter outlines key concepts, syntax, and operational models of co-modeling This approach aims to meet essential criteria such as flexibility, expressivity, scalability, and reusability, empowering ABM practitioners to create intricate coupled models effectively.
Introduction
Chapter 2 outlines the essential requirements for modelers in developing architectures for model coupling, focusing on reusability, scalability, expressivity, and flexibility This thesis introduces a software framework grounded in a conceptual methodology that meets these criteria The framework provides a language for describing specific model couplings, encapsulates existing models, and facilitates the translation across various spatial and temporal scales, all through simple extensions to an existing modeling language The resulting models can be executed and simulated like standard models, featuring dynamic capabilities such as the addition, removal, and on-the-fly replacement of sub-models.
Basic concepts
Co-model
A co-model is essentially an agent-based model where certain agents embody other models, enabling a recursive description akin to MADKIT or DEVS This structure facilitates a unique coupling between the models, enhancing their interactivity and complexity.
Micro-model agents facilitate the description of interactions and collaborations within agent-based models (ABM) This method capitalizes on the versatility of ABM, enabling the reuse of its neutral agent implementation to support various modeling formalisms, which are seen as distinct behavioral architectures A key requirement for these micro-models is the explicit definition of their spatial and temporal scales, ensuring accessibility and modifiability by other micro-models or the overarching co-model they are part of.
Figure 3.1: Co-models extend the base concepts of agent-based modeling formalism by allow- ing agents to be models themselves
Figure 3.1 illustrates a simplified version of the proposal, highlighting that a co-model consists of a group of agents, including those that function as micro-models Conceptually, there is no distinction between standard agents and micro-models within this framework.
Micro-model
A micro-model functions as an instantiated agent within a co-model, serving both as a 'wrapper' that provides access to its attributes and behaviors, and as a 'representative' of the model within the co-model In our implementation, this agent is essentially an instance of the original model Furthermore, micro-models can also act as co-models, enabling a recursive structure that allows modelers to repeatedly define the concept of "models as agents."
Figure 3.2: Micro-model can be co-models too
Integration in the GAMA platform
Why GAMA?
For a proposal on model coupling to be effective, it must incorporate at least one implementation within a recognized programming or modeling environment actively utilized by modelers; otherwise, it risks remaining merely a theoretical exercise.
Kravari's survey evaluates 24 modeling platforms based on criteria such as usability, operational capabilities, pragmatics, security management, and application domains Notably, GAMA stands out as the most versatile platform, currently utilized across a diverse range of fields, with internal estimates from developers suggesting a user base of approximately [insert user base number] This versatility indicates its potential for supporting interdisciplinary projects effectively.
The platform accommodates 2000 to 3000 users and offers robust support for data-driven models It enables the creation of agents from a variety of data sources, including spatial databases, and facilitates the simulation of large-scale models featuring millions of agents with intricate behaviors.
GAMA is an open-source platform designed with a user-friendly high-level agent-oriented language (GAML), featuring a straightforward syntax that caters to modelers who struggle with traditional programming languages like Java or C++ Its extensibility allows users to develop custom features in Java, ensuring that specific needs can be met even if existing functionalities are lacking These attributes make GAMA an excellent choice for implementing and testing model coupling architectures.
Implementation
My approach focused on enhancing the GAML language in two key ways: first, by enabling models to reference other models as micro-models instead of standard imported models, and second, by allowing these references to be instantiated as agents While this may appear to be a minor enhancement, it involved significant behind-the-scenes work, which is elaborated upon in the following sections.
I modified the existing GAMA meta-model to enhance its functionality by integrating a multi-level extension that allows agents to be composed of other agents This involved clarifying the status of models and simulations, enabling models to inherit from regular agents, thus granting them a status akin to species in GAML and making their instances, known as simulations, regular agents Additionally, I refined the role of experiments within the meta-model, ensuring they serve as mandatory access points to imported models to maintain encapsulation As experiments are also agents, they represent micro-models in co-models, allowing the instantiation of a micro-model through the creation of an instance of its corresponding experiment.
Figure 3.3: The meta-model of GAMA [74]
The recent modifications in GAMA (version 1.7) allow co-models to be represented as regular models in GAML, encapsulating multiple micro-models This innovative wrapping method employs a unique importation technique, assigning model identifiers that facilitate access to defined experiments Consequently, any model can define various micro-models that may be instantiated and executed during simulations These significant updates to GAMA's general meta-model required comprehensive integration into the platform, rather than simple extensions or plugins, and are now available to all modelers.
Figure 3.4: Extend a simplified meta-model version of current agent-based modeling platforms to support coupling models (changes are in the dotted rectangle)
Portability
The co-modeling approach is conceptually straightforward, based on the premise that "any model can function as an agent within an agent-based model." However, its operational implementation is more complex, necessitating a multi-level or recursive organization of agents within the target platform or framework, where agents can be composed of other agents.
For effective co-modeling, it is essential to explicitly describe the spatial and temporal specifications of models, ensuring they are accessible to other agents This approach is exemplified in GAMA, where each agent can detail its internal environment and scheduling processes Consequently, the ability to transfer co-modeling concepts to different computing environments heavily relies on the availability of these specifications.
Figure 3.5: Importing model as micro-species features in these environments It should be possible to do it in Madkit, for instance, but not in NetLogo.
Example of use (syntaxes)
Importation
To successfully couple micro-models, the initial step involves identifying and importing them, which requires defining a unique identifier for each model This identifier serves as an alias within the co-model for the imported micro-model An example of this importation process is illustrated in Figure 3.5, where the "import" statement is accompanied by the path to the existing model and its corresponding identifier.
Instantiation
To effectively utilize imported micro-models in simulations, they must first be instantiated using the GAML keyword “create,” which specifies the agent species and may include additional attributes such as the number of agents and their initial properties In GAML, a model serves as a system description, while the actual simulation operations are defined in its associated experiments When integrating micro-models into a co-model's simulation, the modeler must indicate which declared experiments to instantiate For instance, the syntax example in Figure 3.6 illustrates the instantiation of three imported micro-models—“market-grow,” “SaltIntrusion,” and “EnvChange”—with default parameters from the selected experiments “myMarket,” “mySea,” and “myEnv.”
Figure 3.6: Instantiate agents of micro-species
Execution
The multi-level modeling capabilities in GAMA enable the execution of micro-model simulations step by step, allowing modelers to interact with the behaviors and attributes of micro-models effectively Additionally, data exchange between these models is facilitated, providing flexibility in modeling Notably, the spatial and temporal scales of micro-models can be easily accessed through three key accessors, including "shape," which delineates their geographical boundaries Further examples of these data exchange possibilities will be explored in subsequent chapters.
“step” and “starting_date”, which both define its default temporal attributes.
Figure 3.7: Ask agents of micro-species to simulate and exchange data between models
End of chapter
This chapter outlines the key contribution of the thesis: an enhanced ABM meta-model that integrates models as agents, facilitating their inclusion in other models This improved meta-model replaces the previous version in the GAMA modeling and simulation platform, enabling the instantiation and execution of models within models as agents, with minimal modifications to the GAML language The implementation supports full recursion, allowing agents to represent models that are co-models themselves Control and data exchange between micro-models occur naturally through attribute accessors and action executions, including the ubiquitous step action present in all models Consequently, constructing a complex integrated model parallels the process of building a complex agent-based model Additionally, GAMA's ability to handle various formalisms—through GAML extensions or plugins for models and data sources in other languages—positions this enhancement as a valuable foundation for multi-modeling initiatives.
This chapter demonstrates the implementation of a straightforward integrated "Prey-Predator" model utilizing the proposed co-modeling approach Through this practical example, we explore the benefits of co-modeling over traditional methods, particularly highlighting its superior expressivity and reusability.
Objective
This chapter demonstrates the implementation of a practical toy model within the proposed co-modeling approach, aimed at illustrating its application in overcoming the challenges of model coupling discussed in chapter 2 The evaluation of the approach's capabilities, particularly the interoperability of its infrastructure, is conducted based on specific criteria.
• Reusability: Modelers could immediately reuse existing models for build- ing the toy model without too much modifications and efforts.
• Expressivity: The coupling of models could be declared and described explicitly, without any black boxeffect
• Scalability: Models operating by default at different spatial and temporal scales could be coupled together.
• Flexibility: Micro-models could be dynamically added, removed or swapped during the simulation of the co-model.