Key Takeaway
You can calibrate large agent-based transport models inside a single simulation run by updating parameters online, avoiding hundreds or thousands of full-run trials.
ON THIS PAGE
Key Findings
Parameters can be adjusted during the simulation by logging the choice context of agents who replan and optimizing on these recorded batches. That online approach matched observed mode shares in a Switzerland case and converged in about 60 simulation iterations. Tuning only simple constants added negligible runtime; calibrating more parameters took about 2–3 minutes extra per iteration but still avoided repeated full runs. batch optimization.
Data Highlights
1Calibration achieved within a single run, replacing the typical hundreds-to-thousands of full simulation runs used by standard calibration workflows
2Calibration convergence reached in approximately 60 simulation iterations
3Comprehensive parameter optimization added about 2–3 minutes of computation per simulation iteration (negligible overhead when only mode-specific constants were tuned)
What This Means
Engineers and modelers running large-scale transport simulations who need faster, cheaper calibration will benefit—this method reduces the number of full runs and integrates tuning into normal execution. Researchers and technical leads exploring practical agent-based modeling workflows can use the approach to speed up scenario testing and sensitivity analysis. Agent Registry Pattern
Not sure where to start?Get personalized recommendations
Key Figures

Fig 1: Figure 1: Discrete Mode Choice in MATSim

Fig 2: Figure 2: Discrete Mode Choice Calibration Loop

Fig 3: Figure 3: Evolution of calibrated parameters across MATSim iterations

Fig 4: Figure 4: Global mode shares for the simulated modes
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The method depends on a 'quasi-static' assumption: only a small fraction of agents replan each iteration so choice contexts change little between steps; it may break down if replanning rates are high or the system changes rapidly. Results were demonstrated on a 5% sample of a Switzerland scenario and on a specific set of utility parameters, so adapting to different models, larger parameter sets, or other data sources may require tuning. Optimization cost scales with the number of parameters and chosen optimizer; careful selection of objectives and optimizer settings is still needed. Chain-of-Thought Pattern
Full Analysis
The approach embeds parameter calibration inside a single agent-based simulation run by logging the choice variables (travel time, cost, distance, etc.) for agents who replan and performing batch optimization on that recorded data. Because only a small share of agents replan each iteration, the choice context is approximately constant between consecutive iterations (the quasistatic assumption). Using these logged batches, standard optimizers (the authors used a population-based optimizer) update model parameters iteratively so that simulated aggregates—like mode shares and distance-dependent mode distributions—move toward observed targets. Market-Based Coordination Pattern Applied to a 5% national sample for Switzerland, the method jointly calibrated several mode-related utility parameters and produced a good match to census-derived mode shares and distance profiles. Calibration converged in about 60 iterations and, for the full parameter set tested, added roughly 2–3 minutes per simulation iteration; tuning only mode-specific constants produced negligible overhead. The core implication is practical: offline black-box loops (hundreds or thousands of full runs) can be replaced by a single, continuously calibrated run, cutting experiment time and infrastructure needs. Future work includes testing different optimizers, objective functions, and extending the technique to other kinds of parameters such as network capacities. Compound AI System
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Authors have very low h-indices, no affiliations listed and arXiv-only; signals point to emerging work with limited reputation.