A stochastic process approach to agent-based modeling and simulation
Course coordinator: Anna Klabunde
Start: October 19th 2015
End: October 30th 2015
Location: Max Planck Institute for Demographic Research (MPIDR), Rostock
Keynote lecture: Icek Ajzen (University of Massachusetts)
Keynote lecture: Joan Walker (University of California, Berkeley)
Jakub Bijak (University of Southampton)
Jutta Gampe (MPIDR, Rostock)
Anna Klabunde (MPIDR)
Matthias Leuchter (MPIDR)
Matteo Richiardi (University of Oxford, Institute for New Economic Thinking)
Tom A.B. Snijders (University of Groningen and University of Oxford)
Frans Willekens (MPIDR, Rostock)
Agent-based or individual-based models describe how populations evolve, patterns (e.g. social networks) emerge, and collective features (e.g. norms) are established as outcomes of actions and interactions at the micro-level. Simple heuristics or rules govern the actions and interactions. Agents may be humans, institutions or organizations. They have attributes, capacities (e.g. human capital: education and health) and resources (time and capital: physical, financial, social and cultural). Agents are autonomous – they determine their own actions; connected – they interact with each other and with the environment; and adaptive – they act and behave in response to changes in their own characteristics, in that of other agents or in the environment. Feedback and memory are key concepts. Actions and interactions are often outcomes of choice, but are almost always affected by chance, too.
This is an advanced course. We want to help spread important innovations in agent-based modeling and place the methodology on a more stable probabilistic and choice-theoretical grounding. The course will rest on three pillars:
(a)Probability theory, especially the theory of stochastic processes. This is based on the fact that the quantile function (inverse distribution function) is the workhorse of all microsimulation models. We understand an agent-based model as an instance of a stochastic (random) process and it is the aim of this course to connect ABM and probability theory more closely than this is usually done. We start from simple stochastic process models that capture random factors and systematic factors, such as individual characteristics. Individual actions and interactions are determined by values drawn randomly from probability distributions.
(b)Decision theory / choice theory / action theory. We extend stochastic process models by introducing behavior in the form of choice. Action is an outcome of random choice processes that include cognition and social interaction.
We reflect on how agent-based models could be more closely connected to existing research on decision theory. We study two decision theories which are broad enough to incorporate a wide range of different decision contexts. The first one is the theory of planned behavior (TPB), introduced in social psychology by Icek Ajzen. The other one is discrete choice theory whose most prominent originators are Daniel McFadden and Moshe Ben-Akiva. Recent contributions to discrete choice theory (hybrid choice models) have many elements in common with the TPB.
(c)Social network theory. Almost all agent-based models have elements of social interaction, frequently modeled as networks. Stochastic actor-based models for social network dynamics are introduced as an example of how agent-based models can aid statistical inference.
We aim at overcoming the artificial divide between microsimulations and agent-based modeling and show that these methodologies are derived from common ancestors and use a common set of tools from mathematics, statistics and computer science.
Much more so than in other courses on ABM, the topics of calibration, estimation and validation of agent-based models, as well as sensitivity analysis and the design of computer experiments, are treated extensively.
This is a course for PhD students, PostDocs and senior researchers who are currently involved in an agent-based modeling project. We expect participants to have some background knowledge already and to already have decided on a programming language or modeling platform of choice, as this will not be covered in the course.
The course offers four important skills that help to produce agent-based models which are innovative, insightful and statistically sound:
Knowledge in probability theory. We provide an introduction to stochastic processes and convolution. This enables you to start out from truly simple, analytically solvable models whose complexity can then be raised gradually.
Tools for statistical inference in and through agent-based models. It is possible to estimate parameters in agent-based models for example via simulated minimum distance and Bayesian methods, as well as to use agent-based models to identify behavioral rules which are likely data generating mechanisms for real data sets.
Skills in modeling choice behavior. You will be able to connect behavioral theories from fields such as psychology and economics to their mathematical implementations. You will learn about different ways of modeling choice so that you can choose the one adequate for your model.
Design of computer experiments, calibration and validation. You will learn to efficiently explore the parameter space in order to thoroughly understand model behavior, to conduct a systematic sensitivity analysis, and to calibrate a model to data.
The course consists of six days, each dedicated to a different topic, with two days for self-study and assignments in-between. Lectures are in the morning. Afternoons are for computer tutorials. The course will be followed by a three-day workshop providing participants and other researchers an opportunity to present and discuss their work. On the first day of the course, professor Icek Ajzen will deliver a keynote address entitled “The Theory of Planned Behavior: a unifying framework to predict behavior.”
The workshop is co-organized by the IUSSP Panel on Microsimulation and Agent-based Modeling. The subject of the workshop is “The science of choice.” Professor Joan Walker will deliver a keynote address entitled “Modeling choice.” It will cover current innovations in discrete choice models. A separate call for papers for the workshop has been issued. Course participants should participate actively in the workshop.
Lectures and computer lab
The lectures are generally from 9:00 – 12:00; computer labs are from 14:00 to 17:00.
Monday, October 19
9.00-10.00 Introduction to computers and network
10.00-12.00 Lecture: Introduction to stochastic processes (Frans Willekens)
13.00-15.30 Computer lab: Introduction to stochastic processes (Frans Willekens)
Outline and Reading List (PDF File, 180 kB)
16.00-17.30 Süßmilch lecture: The Theory of Planned Behavior: a unifying framework to predict behavior (Icek Ajzen)
Tuesday, October 20
9.00 – 12.00 Lecture and computer lab: Event-history models (Jutta Gampe)
14.00 - 17.00 Lecture and computer lab: Event-history models (Jutta Gampe)
Wednesday, October 21
9.30-11.00 Süßmilch lecture: Studying social influence in networks using the stochastic actor-oriented model (Tom A.B. Snijders)
11.30-13.00 Lecture: Actor-oriented dynamic social network modeling, part II (Tom A.B. Snijders)
14.00-17.00 Computer lab: SIENA package for the statistical analysis of longitudinal social network data (Tom A.B. Snijders)
Outline and Reading List (PDF File, 99 kB)
Thursday, October 22
9.00-12.00 Lecture: Agent-based modeling in continuous time – the example of international migration (Anna Klabunde)
14.00-17.00 Computer lab: Agent-based modeling in continuous time – the example of international migration (Anna Klabunde and Matthias Leuchter)
Outline and Reading List (PDF File, 274 kB)
Friday, October 23
9.00-12.00 Lecture: Calibration, estimation and validation of agent-based models (Matteo Richiardi)
14.00-17.00 Computer lab: Calibration, estimation and validation of agent-based models (Matteo Richiardi)
Outline and Reading List (PDF File, 81 kB)
Monday, October 26
9.30-11.00 Süßmilch lecture: Agent-based modelling in demography: Epistemological and methodological challenges (Jakub Bijak)
Outline and Reading List (PDF File, 67 kB)
11.30-13.00 Lecture: Design and analysis of computer experiments (Jakub Bijak)
14.00-17.00 Computer lab: Implementation and analysis of computer experiments (Jason Hilton)
Tuesday, October 27
Self-study and assignments
Wednesday, October 28
Keynote lecture: Modeling Choice (Joan Walker)
Thursday, October 29
Second day of workshop
Friday, October 30
Third day of workshop
Course prerequisites are (a) experience with agent-based modeling or the successful completion of an introductory course on agent-based modeling, (b) the knowledge of one programming language or agent-based modeling platform, and (c) a working knowledge of R.
If you never used R in your research work, please make sure you have a sufficient knowledge before the course starts, e.g. by attending a free online course such as https://www.coursera.org/course/rprog. Alternatively or additionally you can also use the tutorial website from UCLA (http://www.ats.ucla.edu/stat/r/) or any other R-tutorial which goes into sufficient detail.
Evaluation will be based on active participation in the lectures, computer lab, and workshop, and on successful completion of the required assignments.
General Readings (PDF File, 59 kB)
There is no tuition fee for this course. Students are expected to pay their own transportation and living costs. However, a limited number of scholarships are available on a competitive basis for outstanding candidates and for those applicants who might otherwise not be able to come.
Recruitment of students
Applicants should either be enrolled in a PhD program (those well on their way to completion will be favored) or have received their PhD.
A maximum of 15 students will be admitted.
The selection will be made by the MPIDR based on the applicants’ scientific qualifications.
How to apply
Application is no longer possible.
The science of choice - How to model the decision-making process? Workshop at the MPIDR, October 28-30 - Final Program (PDF File, 389 kB)