A General Cognitive Architecture for Agent-Based Modeling in Artificial Societies

Agent-based artificial societies are a standard tool for studying things like demography, traffic, epidemics and economics. But there is no agreed way to model an agent's decision-making, so every project builds its own, and models rarely get reused or integrated.

• Reviews four families of existing architectures: production-rule systems (cheap, but limited to simple behavior), BDI models (beliefs, desires, intentions) and the BOID variant that adds obligations, normative models where social norms come from outside the agent, and psychology-inspired ones like SOAR, ACT-R and CLARION (rich, but too expensive for thousands of agents).
• Proposes a modular architecture: perception and actuation, learning, working and long-term memory, norms, emotion and personality, physical and social state, reasoning, motivation and attention, planning, and interaction with other agents.
• Runs two simulations that use different subsets of the modules to show one architecture can host different kinds of models.

• Each agent repeats one cycle: observe and talk to others, update memory, learn, reason under the constraints of norms, emotion and physical state, rank motivations by attention, plan, then act.
• Emotion drives short-term, time-pressured decisions; personality drives long-term strategy, and is what makes individuals differ.
• Evacuation (500 agents, two exits): agents pick and switch exits based on what neighbors do, their self-confidence and their fatigue. This short-term scenario used most of the modules.
• Population evolution: Chinese census data, with each agent standing for 10,000 people, modeling births, deaths, and migration driven by wage gaps, distance, and a pull toward home. This long-term, "rational" scenario needed only six modules.

• The authors list the field's open problems: assumptions can be arbitrary, too many free parameters make models hard to falsify, empirical grounding is weak, documentation and replication are poor, and code is rarely reusable.
• Proposed remedies are calibrating behavior with social experiments (like the trust game), fitting micro-level parameters to macro-level statistics, open-sourcing code, and sharing a common knowledge base.
• The two simulations are demonstrations that the architecture works, not benchmarked comparisons against other designs.