Agent-based Modeling and Restructurations Uri Wilensky http ccl

Agent-based Modeling and Restructurations Uri Wilensky http: //ccl. northwestern. edu Center for Connected Learning & Computer. Based Modeling (CCL) Northwestern Institute on Complex Systems (NICO) Northwestern University Indiana University May 21, 2006

Overview The goal of this talk is to argue for a widespread adoption of complex-systems perspectives and methods, and specifically agent-based modeling: In particular, we argue for the use of agent-based modeling – To Reformulate school Content (K - 20) – As experimental methods for evaluating policy – As a method to build and assess theory • Examples of each of these from current CCL work

A thought experiment Imagine a country where everyone uses Roman numerals. The educators in this country were very concerned with problems of numeracy amongst the citizens. • Some focused on numerical misconceptions - If CX is ten more than C, then CIX must be ten more than CI • Some wrote computer-programs to enable students to practice Roman arithmetic • Some construct wooden blocks with X, I, V, C

A thought experiment (cont) Imagine the educators had invented Hindu-Arabic numerals. • Before: the learning gap in arithmetic was immense - only a small number of trained people could do multiplication. • After: multiplication became part of what we can expect everyone to learn.

Restructurations Structurations -- the encoding of the knowledge in a domain as a function of the representational infrastructure used to express the knowledge Restructurations -- A change from one structuration of a domain to another resulting from a change in representational infrastructure (Wilensky, Papert, Sherin, di. Sessa, Kay, Turkle, Noss & Hoyles, 2005; Wilensky & Papert, 2006)

Complex Systems & Restructuration With the invention of complex systems representations such as agent-based modeling, we are now poised to create restructurations of • content for students and for research

Agent-based modeling Creating computer models in which individual computational entities interact to create largescale patterns • The entities are the agents. Each agent has its own descriptive “state variables” (e. g. , age, energy, wealth), graphical depiction and behaviors (simple computational rules e. g. , move, eat, buy) • Out of the interactions of the agents following their rules, emerges a large-scale pattern: the emergent phenomenon

Affordances of Agent-based modeling vs. equational modeling • Agents represent individual elements of the model – The model is built with agents (wolves, molecules, indiv. customers) as opposed to with aggregates (wolf populations, pressure, customer pop. s) • • Agent behaviors can leverage body knowledge Local interactions, Proximate mechanisms. Make use of spatial dimension Model is runnable – visualization of dynamics at multiple levels – Immediate feedback • Model is incrementally changeable – Enabling what-if investigations – Enabling change to model for varying initial conditions • Design Micro-rules that generate macro- pattern • Glass-box models • Requires little formal mathematical machinery

Affordances of Agent-based modeling vs. equational modeling • Non-linearity – Most worldly phenomena are non-linear – Move away from linearity and calculus – Computational reps are non-linear by default • Discreteness – Increasingly discrete models are replacing continuous models

4 areas for agent-based restructuration ripe for the picking • • STEM and social science courses K-20 Policy Theories of learning in social contexts Theories of individual cognition At the CCL, we have begun to work on all four of these. My hope is to encourage others to do so as well.

Agent-based modeling environments • The examples are all implemented in Net. Logo, an agent-based modeling environment developed at the Northwestern CCL Other restructurated education work (particularly of content) has been done with: Agentsheets, Star. Logo, Molecular Workbench (mostly pre-collegiate content) Swarm, Repast, Mason, Ascape (collegiate content and social science research)

Restructurated Content

Connected Chemistry: agentbased molecular chemistry (Levy, Novak & Wilensky) • Example: KMT & Gas Laws – Learned through the exploration of agentbased models – Focus on dynamics of change in addition to traditional curriculum goals, through a complex systems lens • Agents; – Gas particles, operating according to KMT assumptions • Emergent properties – – Pressure, speed distribution, temperature Gas laws Randomness & stability Time lags between perturbation and equilibration

Connected Chemistry: agentbased molecular chemistry (Levy, Stieff & Novak) • Agents: – Particles (molecules) [in gas or solid] • Emergent patterns: – Ideal gas law – Chemical kinetics

Econ. Lab: Agent-based economics (Maroulis) • Example: Oil Cartel – The exploration of the economics of a market with imperfect competition. – Participants experience why cartels are: • difficult to sustain • harmful to consumers • Agents: – Oil producers – Oil consumers • Emergent Properties: – Market price and quantity – Deadweight loss

Prob. Lab (Abrahamson): Agent-Based Prob. & Stats • Agents are computational procedures that make use of a “random” primitive • Emergent pattern is a statistical distribution • Constructing probability by connecting “micro” and “macro” views of randomness • Constructing outcome distribution as a stochastic and multiplicative “transformation” on combinatorial analysis

Material. Sim: Agent-based Materials Science (Blikstein) • Conventional focus: many-to-one (95 variables/18 equations in 30 minutes) • Our focus: one-to-many (simple behaviors that explain a wide variety of phenomena) • Agents: – Atoms • Rule: – atoms “prefer” to be amongst equal neighbors • Phenomena explained: – – – Grain growth Diffusion Phase transformation Solidification Fusion Etc.

Evo. Lab: Agent-based biological evolution (Rand & Novak) How can we facilitate learners understanding of processes that take thousands of lifetimes to occur? By enabling learners to experiment with rules for individual animals or for evolutionary mechanisms and artificially speeding up time, it is much easier to explore “evolutionary space”. Agents: Moths, Wolf, Sheep, DNA, and any Individuals in Ecosystems Emergent Patterns: Camouflage, Natural Selection, Neutral Mutation, Mimicry, Phenotypic Plasticity, Baldwin Effect, Coevolution, and many more

Evo. Lab: Agent-based biological evolution (Rand & Novak) • Agents: – – Competitors Prey Mates Resources • Emergent Patterns: – – – Selective pressures Camouflaging Adaptation of Motion Genetic Drift Bottleneck Effect Baldwin Effect

Cities: Procedural Modeling of Urban Development (Watson & Rand) How do cities grow? Can we use agent-rules to produce quasirealistic city development patterns? Can we introduce some ability to control the outcome? Represent developers, home buyers, and civic government as agents that move around and make decisions. Represent parcels of land as having value dependent on geography and development. Allow learners to paint “honey” and “poison” on to the landscape to influence the decisions of these agents. Speed up the time-scale of the system to allow quick realization of the processes. Emergent patterns: Suburban Sprawl, Road Networks, Central Business District, Zipf’s Law of Urban Population, Clarke’s Rule of Radial Density

NIELS: agent-based electromagnetism (Sengupta) Models depict phenomena in Electrostatics, Electricity, and Magnetism: an emergent perspective n Agents: Electrons, Atoms, Ions Emergent Phenomena: Current, Voltage, Electric field

Educational Policy

School Choice (Maroulis) • Conventional focus: Does choice “work”? • Our focus: Under what conditions would it work or not work. E. g. : • When are “survivors” better than “closers”? • Can we help the market forces along? • Agents: – students, households, schools • Emergent Properties: – Enrollment patterns – Concentration of achievement (Gini ratio)

School Change (Maroulis) • • • Conventional focus: Does a school adopt a reform? Our focus: What are the leverage points for change? Agents: – students, teachers Agent-properties: – e. g. , Closure or brokerage Emergent Properties: – School culture (academic press) – Adoption of innovation – Social capital

Theories of Social Learning

Piaget/Vygotsky (Abrahamson) Agent: marbles player EP: group-learning patterns • ABM for theory of learning – “Runnable” thought experiment – Flexible parametrization – Explicit (proceduralized) – Enables critique/compare (accompanies paper) – Lingua franca for intra/inter-disciplinary discourse To Vygotsky-adjust set best-max-moves-of neighbor End

Theories of individual cognition (Blikstein)

Conservation of Volume (Blikstein) • Conventional focus: When/how do children “get” conservation? • Our focus: Conservation as an emergent result of the behavior and interaction of non-intelligent agents • Agents: – Perceptive elements (detect “height”, “width”, “number”) – Administrative agents: categories of perceptive agents (appearance, history) • Emergent Properties: – Cognitive structures with good performance evolve and survive – The agents are “dumb”, the behavior is intelligent

• • Reasoning about the Rock cycle (Blikstein) Conventional focus: Learning as either a “blackboxed” cognitive activity or a brain science approach Our focus: Learning as an emergent behavior of simpler, easier to understand/model cognitive tasks Agents: – Knowledge retrievers – Knowledge connectors Emergent Properties: – Weak connectors are efficient for short “sentence-sizes” but inefficient for long “sentencesizes” – Strong connectors are inefficient for short “sentence-sizes” but efficient for long “sentence-sizes” connector weathering occurs settles at the bottom of the sea

Summary Table

Center for Connected Learning ccl. northwestern. edu Papers, software, models and curricular units can be downloaded from the CCL web site