EARL I: Budapes t, Hu ngary, 20 07
Actions Across Levels (AAL): A multiple levels perspective on what it means to make sense of complex systems Sharona T. Levy & Uri Wilensky Center for Connected Learning and Computer-Based Modeling University of Haifa Israel
Northwestern University USA
Our question How do we form connections (if at all) between local action and global systemic phenomena? --> Can this be learned?
Goal and context Importance of understanding systems
Learning environments that target or incorporate understanding systems
Variety of ways to understand systems
Developing a framework with which we can explore how people learn or make sense of complex systems
Talk Plan • On complex systems. • Approaches to understanding systems (aggregate, agent-based) • Previous findings: Mid-level construction as an intuitive strategy for understanding systems • AAL: Framework for analyzing reasoning about complex systems • Analysis students’ descriptions of an everyday complex system with AAL • Using the framework: New directions
Complex systems in everyday expressions The rumor spread like wildfire Birds of a feather flock together A system undergoing change
Many individuals
With varying characteristics, considerations and behaviors
Displaying a predictable pattern as a group
Interacting among themselves
… on complex systems
Modeling the forest or modeling the trees?* Aggregate reasoning
Agent-based reasoning
The system is described in terms of populations, stocks and flows, rates of change and feedback.
The system is described in terms of its individual agents, acting and interacting amongst themselves. The collective behavior emerges from these local interactions.
(molecular dynamics; Forrester, 1968)
(Holland, 1995; Bar-Yam, 1997)
The mice (or snake) population increases at a birth rate, which is dependent upon the densities of food and competitors in the environment, and die at a death rate.
A mouse (or a snake) is born, moves around and uses up energy, eats and consumes energy, reproduces, gets older and dies.
The rate at which mice die also depends on the density of the snake population.
A meeting between a snake and a mouse may result in the snake eating the mouse * after Schieritz & Milling, 2003
How do people reason about complex systems? How do we reduce the amount of information to a manageable size? Specifically, information load increases as a result of… …. multiplicity of components …. multiplicity of interactions …. parallel character of actions and interactions …. dynamic nature of system Some things that we do: •
Ignore the dynamics and mechanisms, while focusing on the structures or parts (Hmelo-Silver & Pfeffer, 2004).
•
Work with a “direct” sequential rather than an emergent parallel schema of causality (Chi, 2005)
•
“Copy” from one level to the other (Wilensky & Resnick, 1999)
•
Load control of the system onto a single coordinator, centralized control (Resnick & Wilensky, 1993; Penner, 2000; Jacobson, 2001)
Mid-level construction In previous work*, we have found that when reasoning about ordinary everyday systems, people pervasively employ a strategy we name “mid-level construction”. Students create intermediate level small groups… … in a variety of forms … along one of two trajectories: (a) starting from the agents and grouping; (b) starting from the aggregate and partitioning. The current work is with the same data-set. * Levy, S.T., & Wilensky, U. (in press). Inventing a “Mid-level” to make ends meet: Reasoning between the levels of complexity. Cognition and Instruction.
Experimental Setting •
Sixth-grade class who were engaged with participatory simulations, geared towards understanding emergent phenomena, such as the spread of disease (diverse Chicago Public School, 63% low income)
•
Interviews with ten children, randomly selected out of the class to represent gender and academic success
•
Interview items: Scatter, Rumor, Disease and more
Scatter
Goal:
Scatter interview eliciting potential reasoning about ordinary emergent phenomena
At the beginning of a Physical Education class, the students are standing close together. The teacher tells the students to scatter so they may perform calisthenics. What happens? Can you describe and explain? Supports: • Enrich level • Introduce other level • Connect between levels • Model with tools - pennies, paper and pens
“Scatter” - is that a challenge?.. •
Agent rule: if you’re too close to someone, move away until you find a place in which you can move your arms freely.
• Aggregate level: outbound flow (described thru properties such as: group contour, density gradient, average distance from center, mobility, time till system settles down) • Agent behavior: Agents move away from higher local densities, not necessarily away from the center, resulting in a jagged unpredictable path. •
Emergent behaviors: clustering while spreading, parallel motion of the agents means that two agents can head for the same “spot”. [ row formation ]
Credit: Levy, Abrahamson, Bezold & Wilensky
Simon: Small groups break apart from a clump
Interviewer: Let me bring you these pennies that might help you to explain… I'm a bit confused. Simon: These will be all the people. This one here will be all the room. So they can, some people can move right here some can move right here, some can move right here, some can move right here, some can stay in the middle.
Sean: Staggering outsides-first Interviewer: Who moves? Sean: The outside because otherwise they would have to like push and… And then the inside guys can move around too. Once there's room.
Rachel: Clustering groups on the background of overall spread “There's a bunch of friends talking, and another bunch… “ “.. and then they'd start moving in separate directions so you know they won't be grouped up anymore. And people will probably most likely bump into each other… I mean you run and you look for an empty place... And like even if two people would see the same place, you know, they'd end up, some, one person would end up going in different directions.” “Most of the people will be spread apart but you'll see some groups still together. Like these three will still be together and these three will still be together….”
Faith: Emergent rows
It is kind of hard because people, just they just, line up. And when you look at it after a while, you just see them in lines. It's kind of hard because they're not purposefully trying to make a line, but they just make it.
In 15-18 minutes of describing “Scatter”..
..more about mid-levels •
ALL the children invented mid-levels, half of them more than once [ number of mid-level forms M=1.6, SD=0.7 ].
• We coded the whole interview for the following ideas (Jacobson, 2001): distinction among levels, decentralized control, unpredictability, equilibration processes, nonlinearity. A score for the number of complexity ideas was formed. • Mid-levels are formed either by collecting agents into groups or by subdividing the aggregate, usually not both. • Source-level is associated with the number of expressed complexity ideas [ Spearman’s rho = 0.808, p< 0.05 ]. Students who formed the small groups by collecting interacting individuals expressed more complexity ideas. •
There are ~ three objects in the mid-level [ M=3.3, SD=1.6, range 2-5 ]
•
Core to reasoning about group spread [ 12/16 ]
• Mid-level construction is manifested in a wide variety of forms.
Question
Agent-based reasoning
Aggregate reasoning
We have found two strategies, associated with agent-based and aggregate reasoning, for connecting between levels of description. How are these strategies related to the students’ overall reasoning about the system?
Actions Across Levels (AAL)
* Agent-aggregate complementarity (Wilensky & Stroup, 2003)
Coding with AAL
Sample of analysis A student’s ideas were coded as thinking actions with the AAL framework, counted (removing duplicates) and converted to proportions of the total number of ideas.
AAL profiles
Agent chain i n g Aggregate chaini n g AA chain i n g Paralleli n g Agent rule-ma k i n g Aggregate rule-maki n g AA rule-ma k i n g
Mea n 6 8 11 17 27 16 14
SD 6 5 10 12 8 13 8
! Agent rule-making is predominantly strong ! Chaining, both agent and aggregate, are the least frequent. ! The rest of the actions have a higher degree of variability.
Looking at levels
Level agent aggrega t e AA
Mean 51 24 25
SD 12 15 12
As a whole, students explain the system mainly from an individual’s perspective.
Looking at actions
Action rule-ma k i n g parallelin g chain i n g
Mean 57 17 26
SD 15 12 13
Rule-making > Chaining > Paralleling
Mid-levels are related to strengths Proportions were sorted into “stronger”, “strong”, “less strong” (by ranking and dividing into 3 equal groups). Action Mid-level
N
Agent Aggregate chaining chaining
AA chaining
Paralleling
Agent rulemaking
Aggregate rulemaking
AA rulemaking
Clustering
5
9
7
16
17
26
12
13
Groups
3
9
5
10
28
22
12
15
Outsides-first
3
5
11
4
7
34
27
11
Specific strengths are related to the kind of mid-levels that are formed: – Strong paralleling --> mid-level groups move out all at once. – Strong aggregate rule-making --> mid-level groups move out one at a time. – Differences among “clustering” and “groups” are less clear, but may be related to a relative strength of the first in chaining (32% vs. 24%).
Discussion • Learnability claim regarding agent-based reasoning: – Descriptions at the agent level were more frequent. – However, some students found it difficult to connect between the levels. – To address the strengths of both groups of students, the two forms, agent-based & aggregate reasoning, may both need to be presented, accessible as starting points and bridged. • Students’ modeling of the system based on AAL shows us where more support is needed: – Paralleling: Following parallel actions and interactions – Chaining: Noticing the patterns of evolution among agents and population
Where next? • The AAL framework for exploring people’s reasoning about emergent phenomena: – More contexts – Developmental strands – Chains of actions – Interaction with specific learning environments and supports
AAL strengths Proportions were sorted into “stronger”, “strong”, “less strong” (ranking and dividing into 3 groups). Agent Aggregate chaining chaining
Student
AA Paralleling chaining
Agent Aggregate rulerulemaking making
AA rulemaking
Mid-level
1
4
4
4
22
35
13
17
Clustering, Rows
2
9
9
0
9
27
18
27
Outsides first
3
21
4
8
17
29
13
8
Clustering, Groups, Rules
4
5
10
15
20
15
15
20
Groups
5
10
15
25
15
20
5
10
Clustering
6
0
13
0
0
38
50
0
Outsides first, Rules
7
6
6
11
22
28
11
17
Clustering
8
6
13
13
13
38
13
6
Outsides-first
9
0
0
8
46
23
8
15
Groups, Rows
10
4
4
30
11
19
19
15
Clustering, Rows
Mean
6
8
11
17
27
16
14
SD
6
5
10
12
8
13
8