Portal:Complex Systems Digital Campus/E-Department on Multi-level Governance

Portal:Complex_Systems_Digital_Campus/E-Department_on_Multi-level Governance Multi-level Governance, Prevention and Resilience

Introduction When acting on a complex system, institutions in charge of its governance firstly face the problem of defining desired objectives. Often, these objectives must integrate the conflicting interests and points of view of diverse stakeholders at multiple scales. Then, in order to compromise and to decide on policy actions to match the objectives, it is necessary to build an appropriate understanding of the phenomena, often through modeling, and which includes the effect of the potential actions. (Here, we touch again on the general problem of modeling and reconstructing dynamics from data, addressed in another part of the roadmap). Unfortunately, current methods for addressing action policies (reinforcement learning, viability, etc.) are only practically usable for models in state spaces of low dimensionality. Solutions can be sought in two directions: either by extending these methods to multiscale and higher dimensionality dynamics and multi-level actions (e.g. central and decentralized), or by projecting multiscale dynamics in smaller spaces. The use of stylized dynamics, when possible, is another research direction that could open new possibilities for managing good policy actions on complex dynamics. Finally, dynamics are often uncertain and partially unknown, which implies a difficult compromise between exploitation of better known parts of the dynamics and exploration of worse known parts. This problem can be extended to the reformulation of the problem (including the objectives). This framework similarly addresses problems of control and of design. Main challenges •	Extending the scope of optimal control •	Projecting complex dynamics into spaces of smaller dimension •	Projecting optimal control into high and multiscale dimension space •	Extending exploration / exploitation compromise to problem reformulation •	Co-adaptation of governance and stakeholders’ objectives 1. Extending the scope of optimal control Current methods of optimal control can deal with uncertain non-linear dynamics, and with flexible definitions of the objectives (in viability theory, for instance), but they are limited by the curse of dimensionality: these methods must sample the state space with a given precision, and this requires an exponential computational power function of the dimensionality of the state space. Extending these methods to spaces of larger dimensions is therefore crucial to enable their use in the context of complex systems. One potential approach for addressing these questions is to develop weaker frameworks than the optimal control. For instance, the concepts of resilience and viability could provide interesting sources of inspiration in this respect. These concepts require that the system maintains some important functional properties, which is weaker than the traditional objective of optimal control, which aims to maximise a function. Finally, in some cases, mixing mathematical optimisation of action policies and participatory approaches within an iterative dialog could provide a good compromise between flexibility, social acceptability and rationality. Such approaches would require a specific methodological focus on how to define parts of the problem which can be treated automatically, and how to integrate the results of these optimising algorithms efficiently with other sources within the process of a group decision. 2. Projecting complex dynamics into spaces of smaller dimension (Sara Franceschelli, Pablo Jensen, Ovidiu Radulescu)

Another possibility to tackle the limits of current control methods is to reduce the dimensionality of complex dynamics (for instance, through the identification of slower parts of the dynamics, the aggregation of the state space, the definition of stylized dynamics and so on). This type of work is also very important in negotiation and formulation processes, in order to give stakeholders intelligible materials from which they can easily express their views. We do not know of reduction approaches directed at the local views of the different stakeholders: such approaches would be very interesting. Dimensionality reduction applies to both data (information) and models. Statistical techniques based on Principal Component Analysis determine a linear space containing the essential information. They do not apply to non-linear correlation, when projection should lead to curved manifolds. New methods are needed to cover this case as well. Non-linear Independent Components represent one possible direction of research. Classical model reduction techniques are based on separation of time and space scales. We can cite averaging, singular perturbations, calculation of invariant manifolds. These methods are currently used for applications in physics and chemistry and they should be adapted to take into account the specificities of other domains. Furthermore, complex models are only partially specified. For instance, models in biology are qualitative and knowledge of parameters is only partial. Classical model reduction methods start from models that are completely specified (all parameters are known). There is a need for model reduction techniques that can replace numerical information by ordinal information (one parameter is much smaller than others) or other types of qualitative information. 3. Projecting optimal control into high and multiscale dimensional space (Jean-Christophe Aude, Valérie Dagrain, Guillaume Deffuant, Maud Loireau, Eric Sanchis)

Another possibility is to enlarge optimal control (and any extension beyong optimal control) to high dimensional, multiscale systems. This enlargement should take into account possible distributed actions at different levels, particularly when they are decentralized. Even if the effect of the controls is perfectly known, this perspective is particularly challenging, because in addition to the multi-dimensionality of the system, the control is also in high dimension, with potentially non-linear effects of control coordination. The research should therefore develop new approaches to advance in this direction. The scope of this approach could be extended to the case where even multiple objectives are defined at different scales. The underlying idea behind this proposition is to introduce the concept of "complex objective". It would probably require the introduction of new formalisms to describe the architecture and links between these multiscale objectives. Since they are described at different levels, current control methods are not suitable for tackling this concept. New research should therefore be undertaken in this field using either centralized or distributed control. The latter method is appealing since it allows to take into account different semantics of control and to act at different scales. This concept raises several questions, including: how to couple and synchronise controllers; how to deal with simultaneous and opposite actions on the system; how to handle the different hierarchical levels; how to mix participation/decision making/optimisation; how to make distributed control with a single global objectives or multiple local objectives or both; how to project the results on multiple perspectives. 4. Extending the exploration/exploitation trade-off to governance analysis Decision-makers often face panels of opportunity of actions, among which they have to choose the ones to which they allocate their resources. The outcome and potential success of these opportunities, given their objectives, are often imperfectly known and difficult to evaluate. Therefore, they regularly face a trade-off between exploration of the different available opportunities of action and the exploitation of certain selected opportunities. This trade-off requires experiments at appropriate scales in time and space, and therefore the expenditure of resources, to obtain better knowledge of the value of these opportunities. These expenses must be compared with the potential benefit of such exploration, compared with the mere exploitation of known routines. In the framework of governance, exploration is necessarily made at a given scale of time and space, whereas governance initiatives are performed within open systems and therefore at several scales of space and time. The challenge is thus to propose methods and tools that can go beyond constraints of exploration and bridge the gap between the results of exploratory experiments and full-scale in vivo implementation of governance actions. These methods would have to take into account the reactive and adaptable nature of the targeted systems, as specified in challenge 1. 5. Co-adaptation of governance and stakeholders’ objectives In a multi-level context, identifying the stakeholders and territory concerned is a problem in itself. The main problem is to take into account a variety of objectives related to different scales of time and space, which can be more or less reconciled. The co-existence of different objectives as well as their corresponding potential blueprint, possibly in conflict or in concurrency, raises problems for the management or regulation of the system. Moreover, in some circumstances, the fact that these objectives may evolve according to the evolution of the environment (social context) or may adapt to a dynamical context (Ambient Intelligence) makes the system even more complex to manage or design. The definition of one or several objectives that do not pre-exist and that may emerge is also a key point. We can focus on two challenges:

5.1 The static dimension: governance in the context of heterogeneity of stakeholders, points of view and interests

The challenge is to develop models and methodologies to take into account the large heterogeneity of stakeholders’ viewpoints and interests, which is reinforced by the entanglement of a large range of space and time scales. Multi-criteria analysis is a starting point for solving these problems, but it must be enlarged in order to incorporate several objectives in parallel, and to include the reformulation process. Moreover, the choice of indicators linked to the objectives or their achievement must include stakeholder participation and must be easy to use. On the other hand, the theoretical consequences of their choice, and particularly the potential biases they may introduce, must be carefully investigated. These tools and methods should also propose criteria to analyse the achievement of the objectives or adequacy to the objectives (any-time evaluation).

5.2 The dynamical dimension: evolution of stakeholders' objectives and viewpoints in the governance process

The challenge is to develop models and methodologies to take into account the feedback loops associated with the self-regulation mechanisms attached to the objectives, as well as the interdependance of particular interests during the governance process. For example, by changing the interaction process, the objectives may change the representations, hence the problem and objective formulation and in return the interaction process. This process becomes even more complex in social settings, where multiple objectives and their coordination occur in the same way, but at the collective level. The time scales of model formation, decision-making and the interaction process itself have to be taken into account. These aspects of the problem deal with the question of governance, and renew the participative context where co-learning becomes as important as collective negotiation and decision-making. Moreover, the results of the interaction during the governance process can lead to new views of the problem, and possibly new governance objectives (taking into account, for instance, social acceptability) or new structures in the multiscale architecture of the governance organizations.