Simulated results from an agent-based model examining inequality and innovation in social networks

Moser, Cody1 ; Smaldino, Paul1

Published Nov 14, 2023 on Dryad. https://doi.org/10.5061/dryad.hhmgqnknz

Data files

Nov 14, 2023 version files 430.32 MB

AgentPosition_AgentVars.csv

82.37 MB
AgentPosition_ModelVars.csv

1.14 MB
caveman.csv

32.21 MB
ChangeLinkGini.csv

6.96 MB
ExtendedRuns.csv

1.17 MB
NewGini.csv

13.20 MB
PostCrossover.csv

213.75 MB
random.csv

70.97 MB
README.md

5.97 KB
realworld.csv

385.92 KB
ring.csv

8.17 MB
SuccessRate.csv

819 B

Abstract

Theories of innovation often balance contrasting views that either smart people create smart things or smartly constructed institutions create smart things. While population models have shown factors including population size, connectivity, and agent behavior as crucial for innovation, few have taken the individual-central approach seriously by examining the role individuals play within their groups. To explore how network structures influence not only population-level innovation but also performance among individuals, we studied an agent-based model of the Potions Task, a paradigm developed to test how structure affects a group's ability to solve a difficult exploration task. We explore how size, connectivity, and rates of information sharing in a network influence innovation and how these have an impact on the emergence of inequality in terms of agent contributions. We find, in line with prior work, population size has a positive effect on innovation, but that large and small populations perform similarly per capita; that many small groups outperform fewer large groups; that random changes to structure have few effects on innovation; and that the highest performing agents tend to occupy more central network positions. Moreover, we show that every network factor which facilitates innovation leads to a proportional increase in inequality of performance, creating "genius effects" among otherwise "dumb" agents in both idealized and real-world networks.

README: Innovation-Facilitating Networks Create Inequality

https://doi.org/10.5061/dryad.hhmgqnknz

This is the repository for Moser & Smaldino (2023) "Innovation-Facilitating Networks Create Inequality". The manuscript is publicly available at https://osf.io/preprints/socarxiv/n3hc6.

This repository contains all data analyzed in the study. The Python code for generating the data with an agent-based model, the R script for analysis, and CSVs containing real-world edge lists are in a separate GitHub repository, described below.

For assistance, please contact the corresponding author: Cody Moser (cmoser2@ucmerced.edu)
Recommended citation: Moser, C., & Smaldino, P.E (In press). Innovation-Facilitating Networks Create Inequality. Proceedings of the Royal Society B: Biological Sciences.

Description of the Data

The data are described in terms of what the dataset contains as well as which figures correspond to the respective CSV.
Primary Results
Variables common to these datasets:

NumAgents - Number of agents in the network
ChangeLink - Parameter setting for the probability of link alteration
LinkStep - Step after which link changes begin (should be set to 0 in all datasets)
ProbDiff - Parameter setting for the probability of diffusion of new inventions to neighbors
Initial Path Length - Initial path length of the network
Path Length - Final path length of the network, for cases with link changes
Initial Clustering - Initial clustering coefficient of the network
Clustering - Final clustering coefficient of the network, for cases with link changes
Average Score - The average innovation score of agents in the network
Gini - The Gini coefficient of the network
Step - The step at which the simulation ends
Crossover - Denotes whether the network successfully discovered the final innovation
Incomplete - Denotes if the network is an incomplete graph
Iteration - Denotes the iteration of the run, e.g. for 500 runs of the model, iterations should range 0-499

Files
randowm.csv - Random networks with parameter sweeps on population size, edge probability, diffusion, and random link alteration. Ran to a maximum of 1000 steps. Figures 2, 4A, S1, S7, S9; Tables S1, S2. Unique variable: ProbEdge - Parameter setting for connectivity in generating the Erdos-Renyi random networks.
caveman.csv - Caveman networks with parameter sweeps on clique size, clique count, diffusion, and random link alteration. Ran to a maximum of 1000 steps. Figures 4A, S2, S6, S8; Tables S1, S2. Unique variables:

CliqueSize - Parameter setting for the size of cliques in connected caveman networks; CliqueNum - Parameter setting for the number of cliques in connected caveman networks
ring.csv - Ring networks with parameter sweeps on population size, diffusion, and random link alteration. Ran to a maximum of 1000 step. Figures 4A, S3; Tables S1, S2.
realworld.csv - Real-world weighted and unweighted networks. Ran to a maximum of 1000 steps. Figure 4A; Table 1. Unique variable: Network - States which real-world network the graph was generated from
ExtendedRuns.csv- Random networks with parameter sweeps on population size and edge probability, holding diffusion and random link alteration constant. Ran to a maximum of 10,000 steps. Figure 3. Unique variable: ProbEdge - Parameter setting for connectivity in generating the Erdos-Renyi random networks
NewGini.csv & ChangeLinkGini.csv - Random networks with parameter sweeps on population size, edge probability, diffusion, and random link alteration. Ran to a maximum of 1000 steps. Gini coefficient in this dataset is computed based on a potion's rank rather than its score. Figure S4. Unique variable: ProbEdge - Parameter setting for connectivity in generating the Erdos-Renyi random networks

Additional Dynamics:

PostCrossover.csv - Step-wise data from random networks with parameter sweeps on population size and edge probability, holding diffusion and random link alteration constant. Ran to a maximum of 1,000 steps in addition to 100 steps following crossover. Figures 4, S5. Unique variables: ProbEdge - Parameter setting for connectivity in generating the Erdos-Renyi random networks; CrossStep - Number of steps after Crossover has been attained. Should range 0-99.
AgentPosition_ModelVars.csv & AgentPosition_AgentVars.csv- End step data from random networks where, in addition to model performance, agent-level data is measured to assess the centrality of agents and their overall contribution to model performance. Ran to a maximum of 1,000 steps. Figure 5. Unique model-level variables: prob_edge - Parameter setting for connectivity in generating the Erdos-Renyi random networks. Unique agent-level variables:
SuccessRate.csv- Success rates at a 1,000 step maximum in the task based on the number of agents in ring, random, and caveman networks. Table S2. Unique variables: SuccessRate - Percent of networks which successfully discovered the final innovation within 1,000 steps; Network - Type of network

Missing Data Codes:
NA - Not applicable. For instances when the measure was not calculated.
[] - Not calculated. For instances when the measure was not calculated and the simulation terminated. Generally found in incomplete networks, which terminate on the first step.

Code/Software

The Python code for running the agent-based model for generating the data, the R script for analyzing the data in this repository, and the CSVs containing the edge lists of the real-world networks are stored in the GitHub repository below:
https://github.com/cmoserj/Potions-Model

Methods

The data presented here were generated from an agent-based model of cultural innovation.

Each model is comprised of agents assembled as nodes on a network. The principle model dynamic is elaborated through pairs of agents (dyads) combining sets of items beginning from an initial inventory of six that each agent starts with. Each ideal network is unweighted, but several of the real-world networks (chimpanzee, baboon, and Agta hunter-gatherer) are weighted networks.

Items in each agent's inventory are initialized in an array containing two values: the name of the item and the item's score. In order to craft new items, three specific items must be combined between two agents. With the initial set of six items, there are two valid combinations which can be made: a combination of items a1, a2, and a3 or a combination of items b1, b2, and b3. These will form items 1a and 1b, respectively, which can be combined with items from the initial set in order to make further items. Agents select each item based on a probability calculated by dividing each specific item's score by the sum of the scores of all the items in the inventories. Because each novel item discovered is on another "tier" above the set of items used to create it and has a higher score, this creates path dependency in the model (agents are unlikely to go back and use older items in their inventory over new ones). There are four such "tiers" of items which can be discovered and combined and a fifth tier, which is formed by combining each of the two items on the two separate fourth tiers with one another. The specific scores and item combinations are seen in Fig. 1.

Each ideal network has a number of state variables which are manipulated. Random networks are initialized as Erdős–Rényi networks with the number of agents and critical edge probability as initial variables, ring networks are initialized with the number of agents as initial variables, and connected cavemen are initialized with the number of cliques and clique size as initial variables. Common to these network structures are the probability of diffusion (or the probability that each individual neighbor of an individual agent which discovers an item receives a new innovation when the focal agent discovers one) and the probability of link alteration, or the probability that each agent has one of its links removed and a new one added at the end of each step in the model.