Published in CEUS 2025

Predicting Desire Paths: Agent-Based Simulation for Neighbourhood Route Planning

Computer Environment and Urban Systems (CEUS), 2025 — Bossowski, Szandała & Mazurkiewicz

Authors

• Józef Bossowski — Wroclaw University of Science and Technology
• Tomasz Szandała — Wroclaw University of Science and Technology; SUPSI, Lugano
• Jacek Mazurkiewicz — Wroclaw University of Science and Technology

Abstract

This study addresses how informal "desire paths" emerge when urban infrastructure does not align with pedestrian preferences, focusing on small, local neighbourhoods. It introduces an agent-based modelling framework that builds on path optimisation, affordance theory, and two novel models: a weighted preferences model that accounts for varied pedestrian priorities beyond route length and an obstacle avoidance model that achieves 60% accuracy in replicating real-world desire paths.

Key contributions

• Weighted preferences model — accounts for varied pedestrian priorities beyond simple shortest-path routing, balancing affordance against distance
• Obstacle avoidance model — uses Dijkstra gradients to enable strategic navigation around static obstacles (buildings, barriers)
• Affordance-based path formation — tracks surface wear and positive feedback loops that create emergent desire paths
• Empirical validation — tested against real-world desire paths in a Wrocław housing estate with 60% accuracy

Simulation models compared

Model	Description	MDD (fields)	Accuracy
Basic model	Selects highest-affordance visible field, moves toward goal	22.9 (19.6%)	n/a
Path minimisation	Ties broken by shortest total distance to goal	23.2 (19.5%)	n/a
Weighted preferences	S = wa × affordance − wd × Δdistance	15.8–25.5 (varies by weight)	n/a
Obstacle avoidance	Dijkstra gradient + nearest-field selection with obstacle ray-casting	26.6 (real-world)	~60%

MDD = mean extra distance walked versus the shortest possible route, in grid cells (“fields”); lower is better. Accuracy = share of simulated steps within a few cells of the observed real-world path. The two are independent: the obstacle-avoidance model accepts a slightly longer detour (MDD 26.6) to match actual behaviour most closely (~60%). “n/a” = not reported for baseline models.

Evaluation metrics

• Mean Distance Difference (MDD) — average extra distance walked when taking a desire path vs. the shortest possible route
• Sum of Paths (SP) — total area where footprints exceed a wear threshold, indicating path system extent
• Accuracy (A_i) — percentage of simulated steps within radius i fields from observed real-world paths

Beyond the paper

The simulator builds on the Bossowski et al. (2025) model with several engineering improvements the original study did not model:

• H3 hexagonal grid — replaces the square field lattice, removing diagonal/Manhattan bias and giving uniform 6-neighbour adjacency with no corner-cutting through obstacles.
• Agent waves — agents are released in staggered waves so emergent wear accumulates progressively and the flow network stabilises instead of being dominated by a single run's noise.
• Gaussian blur on the friction field — softens hard surface boundaries so agents transition smoothly between terrain tiers.
• Angular penalty — adds a turn-cost term so agents prefer straight, legible routes over jittery zig-zags.
• Decay / update of wear — the positive-feedback loop is time-aware: older traversals decay and recent ones are weighted up, so the path system reflects current demand.
• Stochasticity (Temperature) — route choice is sampled, not greedy, capturing the spread of plausible human paths.
• Obstacle corner-cutting prevention — when a step would cut diagonally across a building corner, the agent detours around the obstacle instead of jumping it, so paths never clip impassable geometry.
• Terrain-aware wear & recovery — dense grass is harder to wear but persists; lawns wear easily but recover fast, so each surface has its own dynamic affordance.
• Collective ABM wear — all agents share one footprint accumulator, so later agents are drawn to earlier trails; the desire-line network emerges from the crowd.
• Real-world OSM data — the friction map is built from actual map geometry (buildings, water, parks), not a synthetic lattice.
• Adjustable H3 resolution (0–15) — the same model scales from a single block to a whole district.
• Walled-off detection — destinations enclosed by impassable terrain are reported rather than silently failing.
• City-scale parallelism — the agent loop shards across workers sharing one atomic footprint, keeping ABM dynamics at city size.