Dynamic Distribution Guidance

• Dynamic Distribution Guidance is a framework that models guidance as a dynamic process integrating propagation, user route selection, and reaction.
• The method uses a distributive learning scheme where the update weight p adjusts perceived costs based on time, space, and update frequency.
• This approach improves routing efficiency and stability by mitigating overreaction and reducing computational demands compared to full-network dynamic assignment.

Dynamic Distribution Guidance denotes a framework in which guidance is modeled as a dynamic process rather than a fixed instruction. In the transportation formulation developed by Wan, Zhang, and Chen, it appears as a general model of Vehicle Route Guidance Systems (VRGS) that explicitly couples information propagation, user route selection, and information reaction, and updates each traveler’s perceived link cost through a distributive learning weight $p$ that depends on time, space, and related propagation variables (Wan et al., 2023). In this formulation, the central object is not merely a route recommendation, but the spatio-temporal mechanism by which information reaches users, is selectively accepted, and is converted into rerouting.

1. Origin in vehicle route guidance systems

Dynamic traffic assignment and vehicle route guidance are presented as long-standing problems in Intelligent Transportation Systems, and the proposed VRGS model is organized around three interacting components: information propagation, user selection, and information reaction (Wan et al., 2023). The model is intended to generalize beyond static traffic assignment by treating information flow itself as part of the routing system.

The architecture distinguishes two service modes. In a descriptive system, users receive information, choose whether to accept it, and then re-route themselves. In a prescriptive system, the VRGS client computes a new route and presents it, after which the user may accept or reject it. This distinction is operationally important because it changes the ordering of user selection and information reaction, even when the same propagation mechanism is used.

A recurring implication of the framework is that route guidance quality cannot be reduced to the accuracy of observed travel times alone. The decisive factor is the interaction between how information spreads, how users respond to it, and how quickly perceived costs are revised.

2. Core architecture: propagation, selection, and reaction

The first module is information propagation. Raw or aggregated link travel-time updates are generated at their origin and then “flood” outward over time and space according to a propagation strategy $F(X)$ (Wan et al., 2023). This makes propagation an endogenous part of the guidance system rather than a background communications assumption.

The second module is user route selection. Each traveler decides whether to accept or ignore a received information update through a context-dependent selectivity probability,

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$

The notation isolates three classes of determinants: user-side factors, service-side factors, and traffic-related factors. The framework therefore does not assume universal compliance with routing advice.

The third module is information reaction. Once a link-cost update is accepted, the user, or the system client in prescriptive routing, re-routes. The paper lists minimum-cost routing, equilibrium-feedback adjustments, and full traffic-assignment methods as typical reaction rules. The architecture is thus modular: different reaction rules may be paired with different propagation functions and different user-selectivity models.

This decomposition clarifies that Dynamic Distribution Guidance is not identical to rerouting under updated link costs. It is a coupled model of how updates are transmitted, filtered by users, and translated into network actions.

3. Distributive learning scheme and the update weight $p$

The mathematical core of the framework is a distributive learning scheme for perceived link costs. If $C^{Per}$ denotes the perceived cost of a link and $C^{new}$ the true current cost, the update rule is

$$C^{Per}_{new} = (1-p)\cdot C^{Per}_{old} + p\cdot C^{new},$$

with

$p = F(X) = F(x,t,f,v,\ldots).$

Here $x$ is the spatial distance from the information’s origin, $t$ is the time elapsed since the measurement, $F(X)$0 is the information-update frequency, and $F(X)$1 is the effective propagation speed (Wan et al., 2023).

This update can be read as a convex blend of old perception and new measurement. The derivation begins from

$F(X)$2

after which renaming $F(X)$3 yields the learning rule above. The weight $F(X)$4 is therefore the effective learning intensity induced by information freshness and spatial proximity.

Two global measures are defined for the propagation function:

$F(X)$5

To be admissible, $F(X)$6 must satisfy two principles. Finite-amplification requires the total influence to remain finite. Phase-inclination requires $F(X)$7 to approximate the true attenuation $F(X)$8 continuously in space and time (Wan et al., 2023). From a functional-analysis standpoint, these principles place inequality constraints on admissible $F(X)$9.

The role of $P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$0 is therefore broader than a tuning constant. It is the quantitative interface between information propagation and route guidance behavior.

4. Propagation functions and algorithmic realization

The paper gives several prototypical forms of the propagation function $P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$1. For static assignment with no propagation,

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$2

For global feedback with a time gap $P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$3,

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$4

For “natural” global feedback,

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$5

Local variants incorporate space as well as time. The local feedback form with spatial radius $P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$6 and time gap is

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$7

The “natural” local feedback in space and time is

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$8

with

$P_{sel} = F_{sel}(X_{user}, X_{serv}, X_{tra}).$9

A full space-time exponential form is

$p$0

These prototypes range from no propagation to fully decaying spatio-temporal influence (Wan et al., 2023).

The computational procedure follows directly from this scheme. For each update event on link $p$1 at measurement time $p$2, the system sets $p$3 to the measured travel time on that link. For each receiving agent or decision node at location $p$4 and local time $p$5, it computes

$p$6

then updates the stored perceived cost through the distributive learning rule. Once all $p$7 values are updated, each traveler may compute $p$8 and, conditional on acceptance, re-route according to a selected reaction rule such as minimum-cost path on $p$9 (Wan et al., 2023).

The algorithmic structure is therefore event-driven and decentralized in spirit. It does not require that every traveler instantaneously inherit the same network state.

5. Relation to static assignment and classical dynamic traffic assignment

The framework is explicitly positioned against two limiting cases. Static traffic assignment—including Wardrop UE and SO—assumes fixed link costs and $C^{Per}$0; no real-time learning occurs (Wan et al., 2023). At the opposite extreme, many classical DTA models assume perfect, instantaneous, global information, which is represented in the summary as $C^{Per}$1 everywhere at origin time.

The paper argues that the global-perfect assumption can induce oscillation or “hunting.” Dynamic Distribution Guidance relaxes that assumption by letting $C^{Per}$2 decay in space and time. A plausible implication is that the method is designed to regulate overreaction by controlling how strongly and how quickly new information displaces prior perceptions.

The authors also note that the distributive learning approach is computationally lighter than repeated full-network DTA, yet more adaptive than static methods. Table 1 in the paper contrasts three reaction strategies—mathematical assignment, equilibrium-feedback, and minimum-cost routing—under different propagation assumptions, and the resulting dynamic performance in mean travel time depends critically on the propagation function $C^{Per}$3 (Wan et al., 2023).

A common misconception is to treat propagation as a secondary implementation detail while attributing performance entirely to the reaction rule. The formulation argues the opposite: propagation design is itself part of the dynamic assignment problem.

6. Advantages, limitations, and optimization agenda

The stated advantages are threefold. First, the model represents the real spatio-temporal decay of traveller memory and information freshness. Second, it avoids overreaction, or “hunting,” inherent in perfect-information DTA. Third, it lowers computation compared to full-network dynamic assignment at every decision epoch (Wan et al., 2023).

Its limitations are equally explicit. The optimal functional form $C^{Per}$4 and its parameters, including $C^{Per}$5 and $C^{Per}$6, are not prescribed; they must be calibrated via field tests or simulation. The model also abstracts away user-behavior complexity into the selectivity probability $C^{Per}$7, which itself requires empirical estimation. These are not peripheral issues, because both propagation and acceptance determine the realized routing dynamics.

The optimization agenda proposed in the paper has three parts. One is joint optimization of the form of $C^{Per}$8 and its parameters under a chosen reaction rule, treating propagation design as part of the DTA problem. Another is large-scale simulation trials that hold user-choice and reaction rules fixed while fine-tuning $C^{Per}$9 parameters to minimize network-wide measures such as average travel time and time-to-equilibrium. A third is field investigation of real human response curves to spatio-temporal information, thereby directly informing the shape of $C^{new}$0 (Wan et al., 2023).

Later literature uses closely related ideas in other domains by replacing a constant guidance strength with a time-varying or state-dependent one. In diffusion models, for example, limited-interval guidance treats the guidance weight as a function $C^{new}$1 or $C^{new}$2 rather than a single constant (Kynkäänniemi et al., 2024); Feedback Guidance uses a state-dependent coefficient $C^{new}$3 based on posterior confidence (Koulischer et al., 6 Jun 2025); and reinforcement-learning approaches learn per-step guidance trajectories $C^{new}$4 as a sequential decision problem (Zhou et al., 8 May 2026). This suggests a broader interpretation of Dynamic Distribution Guidance as a design principle: guidance quality depends on when, where, and with what intensity corrective information is applied, rather than on a uniform control parameter.

In the original transportation setting, however, the defining move is precise: information propagation is elevated to a first-class decision variable and embedded in the learning weight $C^{new}$5. Under that formulation, Dynamic Distribution Guidance is a flexible and distributive model of route guidance whose stability and effectiveness are governed by the propagation function that links information, perception, and rerouting (Wan et al., 2023).