Source-linked AI summary
Continuous Opinions and Discrete Actions in Opinion Dynamics Problems
Andre C. R. Martins
TL;DR
Discrete-action opinion models cannot retain continuous conviction or represent extremism well, motivating a model with binary actions and continuous internal opinions. The paper introduces Bayesian updating in log-odds and applies it to voter and Sznajd interactions, finding persistent domains and increasingly extreme opinions in both. It concludes that these patterns may characterize the CODA rule across interaction types.
Problem
Discrete models lack memory of past opinions and are poorly suited to representing extremist opinions because they restrict opinions to two values.
Method
The paper represents binary actions alongside continuous internal opinions updated through a Bayesian rule, applying the model to voter and Sznajd interactions on square lattices.
Results
Both models produce clear long-lived opinion domains, with opinions becoming very extreme inside domains and, to a lesser extent, at boundaries.
Takeaways & Limitations
Similar outcomes in the voter and Sznajd models suggest that the observed behavior may be characteristic of CODA rules across different agent interactions.
Takeaways & Limitations
The model assumes α = β, treating neither alternative as inherently favored; when α < 0.5, agents are interpreted as contrarians.
Abstract
from arXiv · showhide
A model where agents show discrete behavior regarding their actions, but have continuous opinions that are updated by interacting with other agents is presented. This new updating rule is applied to both the voter and Sznajd models for interaction between neighbors and its consequences are discussed. The appearance of extremists is naturally observed and it seems to be a characteristic of this model.
1. Introduction
The paper addresses limits of discrete opinion models by proposing agents with binary actions but continuous internal opinions. It applies a Bayesian updating rule to voter and Sznajd interactions on square lattices.
- Motivation: Discrete opinion models represent binary choices but do not retain agents’ past opinions.The paper identifies memory as a limitation of these models.
- Motivation: Discrete models also cannot represent extremist opinions because their opinions have only two values.This limits their ability to describe the emergence of extremism.
- Contribution: The proposed model represents actions discretely while maintaining each agent’s opinion as a continuous internal function.This separates observable behavior from the underlying opinion state.
- Approach: A Bayesian updating rule changes continuous opinions according to agents’ assessments of how likely their neighbors are to be correct.The rule is studied in voter and Sznajd models with neighboring interactions on regular square lattices.
2. Continuous Opinions and Discrete Actions
CODA agents make binary choices from continuous probabilities, updating those probabilities through Bayesian evidence from observed neighbors. Log-odds make these updates additive and distinguish public actions from evolving internal opinions.
- Continuous opinions and discrete actions: Agents assign a probability p that one of two alternatives is best, then choose the alternative with the higher probability.Thus, a binary action can arise from a nonbinary internal opinion.
- Bayesian updating: Neighbor actions provide Bayesian likelihoods that update an agent’s probability of alternative A.For an observed neighbor supporting A, P(A|σj = +1) is proportional to pα.
- Bayesian updating: Odds represent the ratio between belief in A and belief in B, eliminating the Bayesian normalizing constant.The posterior odds are formed after observing a neighbor supporting A.
- Log-odds update: Assuming α = β, the log-odds l = ln(O(A)) change additively by ν = ln α after each observation.The equal-likelihood assumption treats neither alternative as inherently favored.
- Log-odds update: Because log-odds are an invertible transform of p, they measure the same probability while simplifying Bayesian updates.The sign of l determines the corresponding binary choice, whereas its magnitude captures opinion strength.
- Model interpretation: The model contains a continuous opinion layer behind the binary spin actions, so it is more than an Ising model with a different spin-update rule.Extreme opinions must therefore be assessed using l rather than only the sign of the action.
3. Voter Model
The CODA voter model updates continuous internal opinions from neighboring discrete actions, producing stable domains alongside slowly changing boundaries and increasingly extreme opinions.
- Voter-model updating: Each voter-model update changes an agent’s internal probability by ν toward the neighbor’s action, rather than simply copying the action.Disagreement makes the internal opinion less extreme while agreement reinforces it.
- Simulation setup: The simulation starts with balanced, non-extreme opinions on a 50 × 50 lattice and uses α = 0.7, giving ν = 0.8473 per interaction.Initial probabilities range from 0.4–0.49 or 0.51–0.6.
- Domain evolution: Reasonably stable opinion domains form, but small changes continue after the domains are established.Configurations correspond to roughly 4, 8, 20, 80, 400, and 800 updates per agent on average.
- Domain evolution: Straight domain walls strengthen over time because agents encounter agreeing neighbors 3/4 of the time, whereas corners have zero average opinion change.The wall’s two sides therefore become increasingly extreme while corners do not show the same reinforcement.
- Extreme opinions: After 2,000,000 updates, extremists roughly 800 steps from flipping are common, corresponding to p = 4.2 · 10^-295 when ν = 0.8473.The distance-to-flip measure remains 800 steps even when α changes the probability represented by that distance.
4. Sznajd Model
The CODA rule is applied to Sznajd interactions, where agreeing neighbors update multiple agents at once. This produces faster domain formation but the same gradual reinforcement and extreme opinions seen in the voter model.
- Sznajd updating: In the Sznajd model, agreeing randomly selected neighbors influence all their other neighbors, updating six agents simultaneously on a two-dimensional lattice.If the selected neighbors disagree, no update occurs.
- Domain formation: Updating multiple agents per iteration produces much faster convergence and domain formation than the corresponding voter-model dynamics.The larger update scope accelerates the emergence of domains.
- Domain evolution: Boundary agents can still change opinions after domains form, but these changes occur very slowly during longer interactions.This reproduces the slow boundary evolution observed in the voter model.
- Opinion distributions: Sznajd opinion distributions behave like the voter model, with very extreme opinions appearing most frequently.The reported similarity suggests these characteristics may be general features of the CODA model.
- Opinion distributions: Within homogeneous interior regions, continued interaction reinforces already extreme opinions to still more extreme values.The same continuing reinforcement occurs in the voter model.
5. Conclusion
The CODA rule separates continuous opinions from binary choices and produces similar behavior in voter and Sznajd interactions. Both models show long-lived opinion domains with increasingly extreme internal opinions.
- CODA represents agents’ internal opinions continuously while their observed actions remain discrete.The rule is proposed as an alternative to opinion-flipping updates without memory.
- Voter and Sznajd models on periodic square lattices produce very similar results under CODA updating.The authors suggest this behavior may persist across different interaction types.
- Both models form clear domains whose opposing opinions may coexist for very long times.Within domains, agents become highly certain that their local choice is best despite divergent voices elsewhere.