Agent Decision Rules and Bounded Rationality

Summary

ABM agents make decisions through a variety of rule-based mechanisms rather than perfect rational optimization. The three papers illustrate a spectrum of approaches: threshold-based activation of behavioral primitives (Ben Said), utility maximization with stochastic logit noise (Karakaya), and simple probability-based adoption rules (Bonabeau). All share the principle that agents use local, imperfect information and bounded cognitive resources.

Overview

A fundamental design choice in any ABM is how agents make decisions. Unlike rational choice models in classical economics that assume agents maximize utility with perfect information, ABM agents typically operate with bounded rationality — they use heuristics, respond to local stimuli, and may not always act in their best interest. This is both a feature (more realistic) and a challenge (more parameters to calibrate).

Main Content

Decision Rule Architectures

The three papers implement three distinct decision architectures:

1. Threshold-Based Rules (Ben Said et al. 2002)

Agents in CUBES respond to external stimuli through behavioral primitives (BP) that activate when stimulus intensity exceeds a threshold:

• Each BP has an inhibiting threshold ($Inh\_Th{r}_{inf}$, $Inh\_Th{r}_{sup}$) and a triggering threshold ($Trig\_Thr$)
• Stimuli below the inhibiting range have no effect
• Stimuli between inhibiting and triggering ranges affect opinions but don’t trigger action
• Stimuli above the triggering threshold activate the BP and influence the agent’s behavior

This creates a nonlinear, step-like response where small stimuli are filtered out — a form of bounded attention.

See Behavioral Primitives and Thresholds for the full state diagram.

2. Utility + Logit Rules (Karakaya et al. 2011)

Agents compute a utility $U_i$ from four components (quality, promotion, WOM, price), then apply a two-stage purchase decision:

1. Threshold check: $U_i>\alpha$ (buying threshold)
2. Stochastic decision: $\text{logit}(U_i)\ge {\tau }_i$ where ${\tau }_i$ is a random draw

Definition: Logit Purchase Rule (Karakaya et al. 2011)

$$\text{if }{U}_i>\alpha \text{ and }\text{logit}(U_i)\ge {\tau }_i,\text{ consumer }i\text{ purchases the product}$$

where $\text{logit}(u)=\frac{1}{1+e^{k(\alpha -u)}}$, $k$ is a smoothing constant, and $\alpha$ is the buying threshold.

This combines rational utility comparison with stochastic noise, capturing the empirical observation that “human beings do not always act rationally and they may not always purchase a product even though it satisfies consumer’s expectations” (Karakaya et al. 2011).

See Logit Purchase Decision Model for full details.

3. Probability-Based Rules (Bonabeau 2002)

In the product adoption model, each agent $k$ adopts with a probability given by the rate of the master equation — the probability of becoming one is equal to $V({\hat{\rho }}_k)$ per time unit:

$$P({\text{adopt}}_k)=V({\hat{\rho }}_k)=\frac{(1+{\theta }^d){\hat{\rho }}_k^d}{{\hat{\rho }}_k^d+{\theta }^d}$$

This is the simplest architecture: a direct mapping from local information (fraction of adopting neighbors) to adoption probability.

Bounded Rationality Features

All three approaches incorporate bounded rationality:

Feature	Ben Said	Karakaya	Bonabeau
Information	Local stimuli only	Neighborhood WOM	Neighbor adoption fraction
Attention	Threshold filtering	Buying threshold $\alpha$	Implicit in $V$ shape
Noise	Threshold variation	Logit randomness	Probabilistic adoption
Memory	Attitude persistence	Promotion smoothing ($\beta$)	None (memoryless)
Learning	Attitude evolution	Utility updates	None

Stochasticity in Agent Decisions

Role of Randomness

Bonabeau (2002) emphasizes that “stochasticity applies to the agents’ behavior. With ABM, one is not opposed to a noise term added almost out of necessity but rather to an ABM where nothing is certain.” Randomness in ABM is a feature representing genuine uncertainty in individual behavior, not just statistical noise around a deterministic trend.

Connections

• The threshold architecture is detailed in Behavioral Primitives and Thresholds
• The utility + logit architecture is formalized in Consumer Utility Function Components and Logit Purchase Decision Model
• The adoption probability rule feeds into Product Adoption and Diffusion Models
• All decision rules interact with agent heterogeneity through individually varied parameters

See Also

• Consumer Utility Function Components — the Karakaya utility model
• Behavioral Primitives and Thresholds — the CUBES decision mechanism
• Product Adoption and Diffusion Models — the Bonabeau adoption rule
• ABM Methodology and Principles — the broader framework these rules operate within