Variable-Information Guidance

• Variable-information guidance is an adaptive strategy that adjusts recommendations based on dynamic system state and long-term objectives rather than solely local optima.
• It integrates methods like historical simulation, real-time sensor inputs, and off-line optimization (e.g., Particle Swarm) to improve traffic routing, navigation, and parking systems.
• Empirical results show reduced travel times, improved localization accuracy, and lowered control effort, underscoring its practical benefits across diverse applications.

Variable-information guidance denotes a class of adaptive guidance strategies in which the issued recommendation, control, or schedule is not fixed and need not simply reflect the locally or instantaneously best option. In the traffic-management formulation that explicitly defines the term, it refers to a Variable Message Sign (VMS) whose displayed routing advice adapts in real time “not simply to report the objectively faster route, but to drive the network toward a system-level optimum” (Liu et al., 2016). Related work uses analogous constructions in magnetic-anomaly navigation, parking guidance, diffusion sampling, and UAV path following, where the controller or scheduler responds to observability, entropy, occupancy uncertainty, denoising stage, or tracking-error geometry rather than to a single myopic criterion (Ramos et al., 2022, Wu et al., 2016, Chen et al., 23 Jun 2026, Shivam et al., 5 Dec 2025).

1. Conceptual definition and scope

In the VMS setting, variable-information guidance is contrasted with “genuine guidance.” Genuine guidance truthfully reports current travel times on alternative routes, whereas variable-information guidance may display counter-intuitive suggestions, such as strongly recommending Route 1 even when Route 2 is marginally faster, in order to avoid network-level paradoxes and cascading congestion effects. The underlying premise is that a locally truthful signal can be sub-optimal at the network level when route shifts propagate congestion elsewhere (Liu et al., 2016).

Across other domains, the same design pattern reappears in different technical languages. Magnetic-anomaly navigation steers an autonomous vehicle toward regions with stronger observability or larger expected entropy reduction rather than along a shortest path (Ramos et al., 2022). Parking guidance assigns probe cars to spaces and routes that maximize expected information gain of occupancy posteriors rather than only immediate parking success (Wu et al., 2016). Diffusion-model guidance replaces a constant classifier-free guidance scale with a time-dependent schedule optimized for a consistency–coverage trade-off, while UAV path following replaces a constant look-ahead distance with an error-dependent function that enlarges the unsaturated operating region (Chen et al., 23 Jun 2026, Shivam et al., 5 Dec 2025).

Domain	Variable element	Stated objective
VMS route guidance	VMS message $u(t)\in\Omega$	Minimize network-wide average travel time
Magnetic navigation	Yaw-rate choice or guidance law	Minimize localization uncertainty
Parking guidance	Assigned space-and-route action $a\in\mathcal A$	Maximize expected information gain
Diffusion sampling	Time-dependent guidance weights $\{w_t\}$	Control consistency–coverage trade-off
UAV path following	Look-ahead distance $L_0(d)$	Delay saturation and improve path-following

A common implication is that “guidance” is understood operationally: it is a policy that maps state, history, or belief into a recommendation or control input, with the mapping tuned against a system-level criterion rather than a purely local one.

2. System-optimal route guidance via variable message signs

The clearest formalization is the real-time VMS control framework for congestion mitigation. Let $\Phi(q,u,\phi)$ denote the network-wide average travel time per vehicle over the control period, where $q$ is the stochastic vector of system states, $u(t)\in\Omega$ is the VMS message, and $\phi(t)$ are signal timing decisions. The coordinated optimization problem is

$$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$

and, in the VMS-only case,

$$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$

If $a\in\mathcal A$0 vehicles traverse the network, then

$a\in\mathcal A$1

The admissible VMS action set is finite: five messages ranging from “Route 1–strongly” to “Route 2–strongly.” Flow conservation and link capacity are enforced implicitly through the network-loading simulation, while signal timings must satisfy $a\in\mathcal A$2 and $a\in\mathcal A$3 when signals are included (Liu et al., 2016).

The paper instantiates the policy with a Linear Decision Rule (LDR). At each control interval $a\in\mathcal A$4, with observed occupancies $a\in\mathcal A$5, the raw control variable is

$a\in\mathcal A$6

The display decision is then obtained by thresholding: $a\in\mathcal A$7 Here $a\in\mathcal A$8 is the vector of average occupancies $a\in\mathcal A$9 on all 24 links over the previous minute. Off-line training proceeds by collecting $\{w_t\}$0 historical days of $\{w_t\}$1-series, simulating the network for a candidate $\{w_t\}$2, approximating $\{w_t\}$3 by a sample average, and minimizing that objective by Particle Swarm Optimization. In the coordinated scheme, the VMS parameters and signal-control parameters are concatenated into a single decision vector and optimized jointly against the same network-wide objective.

The case study uses a real-world network in Haining, China, with 24 directed links, 4 signalized intersections, and one OD pair $\{w_t\}$4 with two viable parallel routes. The VMS is placed at $\{w_t\}$5 and affects drivers on links 7 and 13 for four OD pairs $\{w_t\}$6. Under compliance $\{w_t\}$7, the reported average travel times over 10 test days are 204.4 s for Genuine VMS + default signals, 198.1 s for LDR VMS + default signals, 186.4 s for Genuine VMS + coordinated signals, and 184.0 s for LDR VMS + coordinated signals. Similar 3–6 s per-vehicle savings are reported under the other compliance scenarios. The stated practical requirements are calibration of driver compliance rates $\{w_t\}$8, reliable real-time link-occupancy estimates, representative historical data for off-line training, and sufficient computational resources for metaheuristic search; once $\{w_t\}$9 is fixed, the on-line computation is described as lightweight enough for 1-min update cycles (Liu et al., 2016).

3. Uncertainty-directed guidance in navigation and parking

In magnetic-anomaly-based navigation, variable-information guidance is explicitly tied to localization uncertainty. One method uses a nonlinear observability matrix

$L_0(d)$0

with local observability Gramian

$L_0(d)$1

The interpretation given is that $L_0(d)$2 quantifies how “rich” the local field gradients are for distinguishing small changes in $L_0(d)$3. Guidance is posed as a receding-horizon dynamic-programming problem over a $L_0(d)$4-step horizon, with action set $L_0(d)$5 and per-step cost

$L_0(d)$6

A second method minimizes expected posterior entropy using the belief $L_0(d)$7 and the one-step expected entropy reduction

$L_0(d)$8

leading to the action-selection cost

$L_0(d)$9

In simulation, observability guidance reduced the trace of the sample covariance by about 20–30% relative to the straight-line case, and in hardware on Turtlebot3 + TwinLeaf MicroSAM it reduced mean localization error from 0.32 m to 0.23 m when $\Phi(q,u,\phi)$0. The entropy-aware method produced a total entropy drop about 20–25% greater than the “blind” shortest-path case over 20 Monte-Carlo runs. The real-time figures reported are $\Phi(q,u,\phi)$1–50 ms for the $\Phi(q,u,\phi)$2 observability-DP at each $\Phi(q,u,\phi)$3 Hz planning update and $\Phi(q,u,\phi)$4 ms for the one-step EER computation with $\Phi(q,u,\phi)$5 on an i5 NUC (Ramos et al., 2022).

Parking guidance uses a Bayesian occupancy model and makes the informational objective explicit. Each space $\Phi(q,u,\phi)$6 is modeled as a Bernoulli variable $\Phi(q,u,\phi)$7, with prior $\Phi(q,u,\phi)$8 and sensor likelihoods

$\Phi(q,u,\phi)$9

$q$0

Beliefs decay toward $q$1 when unobserved: $q$2 For belief vector $q$3 and action $q$4, the reduction in uncertainty is measured by mutual information,

$q$5

equivalently by expected KL divergence. The controller solves

$q$6

subject to feasible route and assignment constraints. The paper identifies this as a single-stage POMDP with reward equal to information gain and uses direct one-step policy search over candidate space-and-route pairs. In a 160-space lot simulator with non-homogeneous Poisson arrivals and exponential parking durations, the near-optimal information-gain policy achieved average relative occupancy-estimation error of about 28% at 10% probe penetration, versus about 50% for a “max-satisfaction” policy and about 45% for random assignment; as probe penetration increased from 10% to 90%, the near-optimal policy’s error decreased only slightly, from about 28% to 20%, and remained more temporally stable than the alternatives. The stated interpretation is an exploration–exploitation trade-off between scanning uncertain spaces and assigning cars to the most likely empty spaces (Wu et al., 2016).

4. Variable information as decoupling, direct dependence, and auxiliary-variable selection

A different but related line of work treats variable information itself as an object to be extracted or selected. For a pair $q$7 with joint density $q$8, the reversible normalization

$q$9

defines a transformed variable that is uniform on $u(t)\in\Omega$0 conditionally on $u(t)\in\Omega$1. Under the stated continuity and monotonicity conditions, the mapping is invertible and yields $u(t)\in\Omega$2, equivalently $u(t)\in\Omega$3. The paper extends this idea to iterative decoupling of $u(t)\in\Omega$4, producing $u(t)\in\Omega$5 such that the transformed variables together contain the same information as the original vector while satisfying $u(t)\in\Omega$6 and $u(t)\in\Omega$7 for $u(t)\in\Omega$8, with Hierarchical Correlation Reconstruction used to approximate the required conditional densities. On this basis it defines direct mutual information through residualized variables $u(t)\in\Omega$9 and $\phi(t)$0, so that

$\phi(t)$1

and proposes a multi-feature Granger-causality framework using residual streams $\phi(t)$2 together with lagged mutual-information curves and feature decompositions of the joint density surface (Duda, 2023).

In incomplete-data analysis, auxiliary variables are selected by an information criterion rather than by direct control. The primary variable is $\phi(t)$3, with $\phi(t)$4 observed and $\phi(t)$5 missing, while an auxiliary vector $\phi(t)$6 is observed in training but not at test time. The joint model is

$\phi(t)$7

The proposed criterion estimates the KL risk for the complete data of primary variables and takes the form

$\phi(t)$8

which is described as an asymptotically unbiased estimator of $\phi(t)$9 up to an additive constant. The paper also states an asymptotic equivalence between this criterion and a leave-one-out cross-validation variant. In the reported simulation, a useful auxiliary variable is selected in over 90% of replicates in the correlated case, while a useless auxiliary is almost never selected for $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$0; in the Wine-data example, the selected auxiliary often reduces the test-set negative log-likelihood of the complete model by several units compared with using $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$1 alone (Imori et al., 2019).

These formulations do not prescribe vehicle or sampler trajectories directly, but they sharpen the informational primitives—residual information, direct dependence, and auxiliary relevance—that many guidance policies attempt to exploit.

5. Variable guidance schedules in path following and diffusion models

In UAV path following, variable-information guidance appears as an exponentially varying look-ahead parameter. With cross-track error $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$2 and heading error $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$3, the proposed look-ahead distance is

$$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$4

where $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$5 and $$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$6. The corresponding line-of-sight magnitude is

$$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$7

and the unsaturated lateral-acceleration command is

$$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$8

The saturation boundary is

$$\min_x E\!\left[\Phi\!\left(q, P_\Omega[f(x,q)], P_{\Omega'}[g(x,q)]\right)\right],$$9

The paper defines the unsaturated region $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$0 and estimates its area by sampling. For $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$1, the reported unsaturated areas are $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$2 and $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$3, with $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$4 percentage points and $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$5. In straight-line simulations, variable $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$6 reduced settling time from 16.2 s to 14.8 s, control effort $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$7 from 210 to 137, and overshoot $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$8 from 40.4 m to 12.7 m. On an elliptical path, it reduced settling time from 52.5 s to 46.8 s, $$\min_x E_q\!\left[\Phi\!\left(q, P_\Omega[f(x,q)]\right)\right].$$9 from 612 to 395, and $a\in\mathcal A$00 from 35.1 m to 14.3 m (Shivam et al., 5 Dec 2025).

In diffusion models, variable guidance is formulated as a time-dependent classifier-free guidance schedule. Given unconditional and conditional scores,

$a\in\mathcal A$01

the guided score is

$a\in\mathcal A$02

The information-theoretic formulation introduces a clean-endpoint tilted reference

$a\in\mathcal A$03

and minimizes a forward KL between the actual CFG-induced endpoint law and that reference. The resulting objective decomposes into a consistency term

$a\in\mathcal A$04

and a coverage cost

$a\in\mathcal A$05

The paper then derives trajectory-level formulas involving $a\in\mathcal A$06, score divergences, and $a\in\mathcal A$07, so that the schedule can be optimized from samples and score evaluations without explicit density estimation. On ImageNet-512 with EDM-XXL, the learned schedules are reported to move the precision–recall frontier outward; for matched average guidance, they achieve FID $a\in\mathcal A$08 versus $a\in\mathcal A$09 for constant CFG, with recall improvements of several percent. On COCO with SD-XL, the learned schedules improve the FID–CLIP trade-off, for example at mean guidance $a\in\mathcal A$10, with CLIP increasing by 0.01 and FID decreasing by 0.4 points relative to constant CFG. The learned schedules are described as allocating low guidance at very high noise, rising in the mid-noise region, and moderate guidance at low noise (Chen et al., 23 Jun 2026).

OUSAC extends the scheduling view to acceleration of diffusion transformers. It replaces fixed-scale CFG with

$a\in\mathcal A$11

and skips the unconditional branch when $a\in\mathcal A$12. The schedule $a\in\mathcal A$13 is optimized against a fidelity term and a sparsity penalty that counts full-CFG steps, using an evolutionary search over real-valued encodings followed by thresholding. A second stage assigns different SVD truncation ranks to different transformer-block regions to preserve caching performance under variable guidance. The reported results are 53% computational savings with 15% quality improvement on DiT-XL/2 at ImageNet $a\in\mathcal A$14, 60% savings with 16.1% improvement on PixArt-$a\in\mathcal A$15 at MSCOCO, and $a\in\mathcal A$16 speedup on FLUX while improving CLIP Score over the 50-step baseline (Sun et al., 16 Dec 2025).

6. Common structure, recurring trade-offs, and limitations

Several recurring design choices cut across these otherwise heterogeneous formulations. First, the controlled object is typically a low-dimensional policy parameterization rather than a free-form control law: an LDR matrix $a\in\mathcal A$17 for VMS and signals, a finite yaw-rate set $a\in\mathcal A$18 with horizon $a\in\mathcal A$19 for magnetic navigation, a finite action set $a\in\mathcal A$20 of space-and-route pairs for parking, per-step guidance weights $a\in\mathcal A$21 for diffusion, and a three-parameter function $a\in\mathcal A$22 for variable $a\in\mathcal A$23 (Liu et al., 2016, Ramos et al., 2022, Wu et al., 2016, Chen et al., 23 Jun 2026, Shivam et al., 5 Dec 2025).

Second, the state signal supplied to the guidance law is explicitly informational. In VMS control it is the recent history of link occupancies; in magnetic navigation it is the local observability Gramian or the expected posterior entropy; in parking it is the posterior occupancy vector $a\in\mathcal A$24; in diffusion it is the reverse-time denoising stage and score-field geometry; in auxiliary-variable selection it is the estimated impact of $a\in\mathcal A$25 on complete-data KL risk (Liu et al., 2016, Ramos et al., 2022, Wu et al., 2016, Chen et al., 23 Jun 2026, Imori et al., 2019).

Third, the literature repeatedly identifies a tension between immediate local preference and system-level or long-horizon performance. In traffic routing, truthful “genuine guidance” may worsen network-wide delay through paradoxes and cascading effects. In parking, maximizing information gain competes with maximizing immediate parking success. In diffusion, stronger CFG improves condition consistency but reduces diversity and coverage. In UAV path following, a small fixed look-ahead can improve near-path responsiveness but drives the vehicle more quickly into turn-rate saturation when errors are large (Liu et al., 2016, Wu et al., 2016, Chen et al., 23 Jun 2026, Shivam et al., 5 Dec 2025).

Fourth, many of the methods rely on computationally heavier off-line optimization followed by lighter on-line execution. The VMS LDR is trained off-line by simulation and Particle Swarm Optimization but then evaluated within 1-min update cycles. The magnetic-navigation DP and one-step EER controllers are reported to fit real-time onboard budgets on an i5 NUC. Adaptive CFG schedules are learned through repeated trajectory sampling and objective estimation, while OUSAC uses an evolutionary search plus a calibration stage for caching (Liu et al., 2016, Ramos et al., 2022, Chen et al., 23 Jun 2026, Sun et al., 16 Dec 2025).

The practical limitations are correspondingly domain-specific but structurally similar. VMS deployment depends on reliable link-occupancy estimation and calibration of compliance rates. Parking guidance depends on calibrated sensor likelihoods and belief-decay modeling. Magnetic navigation assumes a known magnetic map and requires derivative estimation on a discretized field. Auxiliary-variable selection assumes the conditional model for $a\in\mathcal A$26 is at least close to correct. Diffusion-schedule optimization assumes access to conditional and unconditional scores and requires stable divergence estimation, while OUSAC notes that variable guidance patterns undermine standard caching methods designed for constant scales (Liu et al., 2016, Wu et al., 2016, Ramos et al., 2022, Imori et al., 2019, Chen et al., 23 Jun 2026, Sun et al., 16 Dec 2025).

Taken together, these works present variable-information guidance not as a single standardized algorithm, but as a recurring systems principle: guidance is improved when the policy reacts to informative structure in state, belief, history, or trajectory stage, and when optimization targets the relevant system-level criterion rather than a locally myopic surrogate.