The overall concept of the project is based on two main Research Objectives (RO) that are interconnected according to the diagram:
Figure 1 - Overall approach.
- Upper Row: We start with data processing that will be used to create predictive models and to estimate the parameters for massive scale agent-based simulations. The information produced so far will be compressed using manifold learning techniques to reduce visualization burden and explosive dimensionality problems. This output will feed the policy optimization part of the framework.
- Lower Row: We will maintain and act on several different event sequences at once (timeline diagram). And show the results using powerful, interactive spatial representations.
RO1 - Air pollution, health and mobility in the metropolitan area of Barcelona
The goal is to address the urgent need for predictive models that use machine learning and artificial intelligence (AI) methods for automated inference of the effect of health and environmental policies using public data, mathematical models and simulation results. In this pilot study, we will focus mainly on the following Research Activity
RA1.1 - Model development
At first, a common methodological framework will be elaborated to enable a consistent and appropriate scenario definition process. During this process several models will be discussed for mobility, pollution and health in the metropolitan area of Barcelona.
RA1.2 - Define scenario extent and detail for the interaction of pollution and citizens' health
The methodological framework will define required data for mobility, air pollution and health modelling and it will also detail methods and processes of data collection to allow consistent and coherent scenario definition, suitable to further simulations. Capturing the interplay between pollution and health is complicated because both variables results from the aggregation of many small decisions of individual agents, compounded over an extended period of time, plus the effect of the enforced global policies.
RO2 - Methods for massive scale simulation for societal analysis
Massive scale agent-based simulations are instantiated using public open datasets on demographics, health and pollution as starting point and the agents are set free to operate in the context indicated by a pre-existing model of society. The agents need to behave realistically, detecting changes in their environment and adapting accordingly.
This is subdivided in the following Research Activities (RAx)
RA1.1 - Infrastructure for massive agent based simulations for social simulations
IPER will need a fully integrated, scalable bundle that includes software, model, simulations, infrastructure, and interactive tools to explore different implications in decision making processes.
The core of the platform will be split in two parts: frontend and backend services. The frontend will use geospatial databases and tools (PostGIS, CartoDB) and powerful interactive technologies to allow users to explore the content and the results of the models and to estimate the effects of health and environmental policy changes. The backend will be based on advanced clustering solutions including NVIDIA CUDA technologies to run massive scale simulations of societal models. This part continues an existing research line in BMT-UPF so we can plan to reuse as much as possible and adapt it to this new project.
RA1.2 - Learning and adaptation
Reasoning, planning and learning in multi-agent systems is typically addressed in the framework of reinforcement learning and optimal control. This requires defined optimality criterion in the form of a cost function. In multi-agent systems, the agents need to plan with their local view on the world and to coordinate at multiple levels. This is especially true for a complex and multifaceted problem like the interaction between contamination of the cities and people's health. The agents will need to reason about the knowledge, observations and intentions of other agents, which can in turn be cooperative or adversarial. Multi-agent learning algorithms need to deal inherently with non-stationary environments and find valid policies for interacting with the other agents.