Campus Dining Collective Behavior Simulation LLM-Agent × ABM

In a dense campus, cognitive LLM-Agents drive collective time–place–mode–route decisions to quantify how information strategies redistribute queues and road loads and shave peaks at the population level.

Research Background & Value Link to heading

• Population view: unify chain decisions for thousands during peaks, measuring aggregate space use and peak pressure.
• Interpretability: per-decision rationales with provenance enable causal tracing from individuals to the crowd.
• Actionability: compare global vs. beaconed information to quantify reductions in queue incidents and flow variance σ.

Research Methodology Link to heading

1. Decision framework: four loosely coupled choices—departure time, dining destination, travel mode, and route—coordinated end-to-end. Each step carries over short-term memory (STM) and prior-step environment snapshots, emits transparent rationales, and remains traceable at the individual level and aggregable at the population level.

2. Memory & sensing: a sticky memory formed by a long-term preference summary plus recent events incrementally assimilates preferences over distance, price, crowding, and road load. When information strategies are enabled, dining-crowding and road-congestion enter the LLM-Agent as previous-step snapshots (≈5 minutes earlier), serving as contextual priors for the current decision.
3. Prompt–data contract:
   

   Aspect	Meaning	Example prompt phrase (decision)
   Joint trade-offs (generic)	No fixed priority; weigh time, distance/price, preferences, crowding/congestion, and context together	“Consider all signals jointly with no fixed priority.” (all decisions)
   Semantics & provenance (generic)	Crowding/congestion are previous-step snapshots; unknown ≠ zero; provenance must be tagged in reasons	“Crowding/road info are previous-step snapshots; unknown ≠ zero; tag provenance, e.g., [crowd_provenance: global_prev_step] / [road_provenance: beacon_prev_step].” (all)
   Fairness & explainability (generic)	Avoid unsupported identity claims; provide traceable, aggregable rationales	“Fairness: avoid unsupported identity claims; provide transparent reasoning.” (all)
   Salience updates (generic)	Apply small preference adjustments as sticky-memory tuning rather than hard rules	“Apply small salience adjustments in [-0.15, 0.15] as gentle memory updates.” (all)
   Feasibility & time grid (departure)	Choose within the allowed grid under class boundaries; prefer contiguous feasible blocks	“Choose from the allowed time grid under class constraints; prefer a contiguous block around the target length.” (departure)
   Tie-breaking under near-equivalence (departure)	Use the shortlist as a soft tie-breaker; maintain stochastic consistency if needed	“When options are close, use the shortlist as a soft tie-breaker with a consistent stochastic token.” (departure)
   Time pressure & preferences (place)	Balance distance, price, personal preferences, and time pressure; be conservative under clearly high crowding	“Balance distance, price, preferences, and time pressure; be conservative under clearly high crowding unless explicitly justified.” (place)
   Route features & weather (mode)	Trade off distance, route length/congestion/turns, weather, and biking overhead; reduce exposure under high congestion	“Trade off distance, route features (length/congestion/turns), weather, and biking overhead; under high congestion, reduce exposure.” (mode)
   Length–congestion–turns (route)	Jointly optimize total length, congestion (with coverage ratio), and turning complexity; avoid overconfidence under unknowns	“Balance total length, known congestion (and coverage ratio), and turns; avoid overconfidence when congestion is unknown.” (route)

Study Area & Data Link to heading

• Area: The empirical setting is Tongji University’s Siping Road Campus in Shanghai (100.98 ha; 31,653 users), a high-density campus with seven dining facilities and 16 functional zones.
• Data: Data comprise 21 weekdays of Campus Smart Card System logs from May 2023 (~6.14M records; ~27,800 diners with ~250k lunch transactions, 10:00–13:30), supplemented by 198 behavioral-preference surveys capturing dining duration, strategy acceptance, information willingness, and adjustment tolerances.

Usage & Scenarios Link to heading

Usage & Workflow Link to heading

Configure information (global/road/beacons x-y-r) → run collective simulation → inspect map/curves/logs → export evaluation HTML and playback.

Use Cases Link to heading

Embed information interventions and demand-side policies (e.g., staggering, mode/route guidance) into an agent-based behavioral simulation to quantify impacts on congestion exposure, travel-time loss, accessibility, and equity; future extensions will cover supply-side scenarios (e.g., dynamic pricing, pop-up takeaway stations) to enable integrated policy appraisal and the estimation of welfare/elasticity indicators (e.g., VOT, consumer surplus, demand elasticities).

View Sample Report

Core Conclusions Link to heading

Mechanism Finding Link to heading

LLM-Agent log excerpts (table summarizes decisions across information scenarios).