Dynamic Targeting Methods

• Dynamic targeting is an adaptive design pattern where target selection and treatment decisions are updated continuously based on real-time observations and system constraints.
• It is applied across diverse fields such as spacecraft autonomy, online advertising, and biomedical interventions, demonstrating practical uses from retasking sensors to adaptive drug dosing.
• Advanced implementations employ closed-loop sensing, constrained optimization, and graph-based methods to balance responsiveness with stability under tight resource and timing envelopes.

Dynamic targeting is a family of targeting methods in which target selection, treatment assignment, or follow-on action is conditioned on current observations, evolving system state, or explicitly modeled constraints rather than fixed once in advance. Across the cited literatures, the common pattern is a closed loop from signal acquisition to decision and action: lookahead imagery can trigger follow-on nadir imaging in low Earth orbit, per-impression advertising systems can recompute bids from fresh behavioral context, Markovian treatment systems can compare state-aware treatment effects to state-specific thresholds, and target-specific optimization paths can score candidate training data by their loss reduction along a validation-induced trajectory (Chien et al., 22 Aug 2025, Wang et al., 2016, Hu et al., 30 Jun 2025, Yang et al., 10 May 2026).

1. Cross-domain scope and core structure

Dynamic targeting does not denote a single domain-specific algorithm. It appears instead as a recurring design pattern in which a targeting mechanism is updated from newly available information and executed under operational constraints. In spacecraft autonomy, the target is a physical ground location or phenomenon. In advertising and marketing, the target is a user, customer, or customer–offer pair. In dynamic systems and control, the target is a treatment decision conditioned on congestion or capacity state. In data selection, the target is a subset of training examples chosen for a downstream task (Chien et al., 22 Aug 2025, Yang et al., 2023, Shchetkina, 7 Jul 2025, Hu et al., 30 Jun 2025, Yang et al., 10 May 2026).

Domain	Targeting signal	Representative mechanism
Spacecraft autonomy	Lookahead imagery	Onboard analysis, retasking, slew, and focus imaging
Advertising and marketing	Behavior, entities, uplift, or aggregate query feedback	Per-impression bidding, entity-graph reasoning, constrained policy optimization
Markovian treatment systems	State-aware treatment effect	Compare CADE/CATE-like quantities to state-specific thresholds
Social and opinion networks	Network position and contagion dynamics	Seed selection on influence graphs or multiplex networks
Targeted data selection	Validation-induced optimization path	Score candidates by normalized endpoint loss drop

Two contrasts recur. First, dynamic targeting is set against pre-planned or static assignment: the CogniSAT-6 paper contrasts DT with Earth-observation tasking planned hours to days in advance, and the marketing papers contrast it with rule-based, seed-based, or propensity-only targeting (Chien et al., 22 Aug 2025, Abutbul et al., 2023, Li et al., 2024). Second, dynamic targeting usually couples targeting to a hard resource or timing envelope, such as a 60–90 s orbital window, a limited number of differentially private aggregate queries, bounded campaign budgets, or a small set of phase-controlled optical feedback parameters (Chien et al., 22 Aug 2025, Shchetkina, 7 Jul 2025, Lu et al., 2023, Mey et al., 3 Jun 2025).

2. Closed-loop sensing and physical control

In spacecraft autonomy, dynamic targeting is defined as “a spacecraft autonomy concept in which sensor data is acquired and rapidly analyzed and used to drive subsequent observation.” The low Earth orbit realization on CogniSAT-6 uses lookahead imagery 40–50° ahead of nadir, onboard transfer to an Intel Myriad X edge processor, autonomous analysis for clouds, thermal anomalies, or spectral signatures, translation of detections into targeting and instrument settings, a 40–50° along-track slew, and acquisition of a near-nadir “focus” image. The same multispectral imager serves as both lookahead and follow-up sensor, and the full sensing–analysis–repointing loop must fit within a roughly 60–90 s lookahead-to-nadir window; at 45° lookahead the paper states a lead time of about 74 s. The onboard algorithms include CNN inference, Spectral Angle Mapper, matched filters, and spectral unmixing, with simple dynamic range stretching used instead of full radiometric correction, and the architecture can initiate the slew before lookahead analysis completes to save time (Chien et al., 22 Aug 2025).

In semiconductor lasers, dynamic targeting refers to active control of external-cavity modes under optical feedback. The relevant physical model is the Lang–Kobayashi system

$$\dot{E}(t) = \frac{1}{2}(1-i\alpha)N(t)E(t)+ \kappa E(t-\tau)\exp(i \phi),$$

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$

with feedback rate $\kappa$, delay $\tau$, and phase $\phi$. The paper focuses on the maximum gain mode, for which the control relation $\phi = \tau \alpha \kappa$ yields frequency shift $\Delta\omega = -\alpha\kappa$. By dynamically co-adjusting $\kappa$ and $\phi$ along this manifold, the laser can remain on the same external-cavity mode while its emission frequency moves continuously. A free-space experiment demonstrates reliable and reproducible tuning over 2.1 GHz, and simulations suggest tuning ranges of about 120 GHz and scan speeds exceeding $10^{17}$ Hz/s (Mey et al., 3 Jun 2025).

A third physical-control instance is adaptive drug dosing against stochastic resistance. “CelluDose” formulates dosing as a continuous-state, continuous-action MDP with a 7-day horizon and decisions every 4 hours. The state includes phenotype concentrations and the aggregate growth-rate feature

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$0

while the reward combines terminal extinction objectives, cumulative dose penalties, and potential-based shaping from $\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$1. In simulation-trained DDPG policies, both the single-drug and combination-therapy controllers achieved a 100% success rate at suppressing cell proliferation, while maintaining low-dose no-event baselines and reacting aggressively to emergent resistant subpopulations (Engelhardt, 2019).

3. Audience targeting, prospecting, and real-time personalization

In online advertising infrastructure, dynamic targeting is realized at impression granularity. Real-Time Bidding treats each ad impression as a real-time auction in which DSPs typically have about 10–100 ms to transform a bid request into a feature vector, predict response such as CTR or CVR, estimate the bid landscape, decide whether to bid, compute a bid price, and choose a creative. The decision is dynamic because it is recomputed for each user–context instance rather than fixed at campaign launch, and it is further adjusted by pacing, budget control, and feedback from observed clicks and conversions (Wang et al., 2016).

For service targeting without seed users, the EGL system replaces “seed-user → look-alike users” with “service-related entities → entity graph reasoning → users.” Its offline pipeline builds an entity dictionary, extracts 30-day entity sequences, mines an entity graph with the Three-stage Relation Mining Procedure, stores graph data in Geabase, and computes user–entity preference scores

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$2

Online, marketers input service-related entities, expand them to $\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$3 hops on the graph, select entities interactively, retrieve users with high preference scores, and export the top-$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$4 audience. The graph is updated weekly, user–entity preferences daily, and end-to-end targeting jobs complete in 2–4 minutes on average (Yang et al., 2023).

Under third-party privacy constraints, dynamic targeting becomes a sequential query design problem. “Blind Targeting” assumes that advertisers cannot see row-level data and may issue at most $\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$5 aggregate queries over hyperrectangles in feature space, receiving noisy treatment-effect estimates protected by a Gaussian differential privacy mechanism

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$6

The method places a GP prior on the latent CATE surface, performs integral posterior updates for region-level queries, and uses a targeting-aware acquisition function

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$7

to prioritize high-uncertainty regions near the treatment/no-treatment boundary. On the Criteo uplift dataset with roughly 14M users, an intuitive benchmark achieved only 33% of the non-privacy-preserving targeting potential in one regime, while strategic querying achieved 97–101% and was statistically indistinguishable from Causal Forest (Shchetkina, 7 Jul 2025).

In native e-commerce advertising, dynamic-product-ads require targeting not only an audience but also a product within a large catalog. Yahoo Gemini’s DPA work introduces Conversion-Prospecting and Trending-Prospecting. Conversion-Prospecting predicts conversion-given-click and sets the per-user, per-product bid

$\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$8

thereby coupling eligibility and bidding to advertiser CPA goals. Trending-Prospecting instead computes a lookalike tendency score $\dot{N}(t) = \frac{P-N(t)-(1+2N(t))E^2(t)}{T},$9 from advertiser pixel feeds and general-population negatives, then uses advertiser-specific thresholds to trade off scale and quality. Online tests reported roughly 10% DPA-delivery lift and 1.58% DPA-revenue lift for Conversion-Prospecting, and roughly 13% delivery lift and 6.33% revenue lift for Trending-Prospecting, while 84% and 88% of spend, respectively, remained within the paper’s stated CPA “happiness” criteria (Abutbul et al., 2023).

A closely related per-campaign framework is “Segment Discovery,” which combines uplift modeling with constrained optimization for customer interventions. In the online spend-threshold experiment reported there, the production policy generated a 0.36% lift in revenue and a 5.49% lift in completion rate relative to the control policy, while the retention case showed that the best uplift-based policy targeted only about 6.47% of customers and still improved over score-based or blanket targeting (Li et al., 2024).

4. Threshold rules, constrained optimization, and target-aligned learning

A central theoretical result for dynamic targeting in Markovian systems is that optimal policies can retain a simple threshold form even when treatments alter future congestion. In “Optimal Targeting in Dynamic Systems,” units arrive sequentially with exogenous covariates $\kappa$0, system state $\kappa$1, binary treatment $\kappa$2, and outcome $\kappa$3, while the state evolves as a finite-state Markov chain. The paper decomposes the conditional policy gradient into a conditional average direct effect,

$\kappa$4

plus a state-only indirect term $\kappa$5, and proves that an optimal policy has the form

$\kappa$6

with state-specific thresholds $\kappa$7 encoding the cumulative congestion cost of treating an additional unit in state $\kappa$8. The proposed estimator first learns $\kappa$9, then searches over threshold vectors using off-policy evaluation; empirical studies show considerable long-run gains relative to direct CATE targeting rules (Hu et al., 30 Jun 2025).

In large-scale marketing, dynamic targeting often appears as repeated constrained policy optimization rather than sequential control. “Optimizing Scalable Targeted Marketing Policies with Constraints” formulates targeting as a linear program over customer–action assignment probabilities $\tau$0, maximizing estimated incremental profit subject to per-segment volume constraints, weighted outcome constraints, similarity or fairness constraints across segments, and per-customer action limits. The paper adapts PDLP to this setting, showing per-iteration cost $\tau$1 and extending two-loop convergence guarantees to obtain $\tau$2 iterations. In an application with more than 2 million customers and six actions, PDLP solved instances with both volume and similarity constraints that caused primal simplex, dual simplex, and barrier methods to fail (Lu et al., 2023).

Dynamic targeting also appears in machine learning as targeted data selection. “Let the Target Select for Itself” argues that many attribution methods suffer from “reference path bias” because they score examples along optimization trajectories induced by the heterogeneous candidate pool rather than by the target task. The proposed alternative is a validation-induced flow,

$\tau$3

combined with a zero-order score given by relative loss reduction between two validation-warmup checkpoints,

$\tau$4

The same compact target-aligned warmup can be reused across candidate pools, with reported warmup below 1 minute per target, scoring time around 3.3 h across all pools/tasks, and storage below 10 MB in the Llama-3.2-3B experiments (Yang et al., 10 May 2026).

5. Network and graph-based targeting

In multiplex behavioral epidemiology, dynamic targeting is operationalized as selection of initial pro-vaccine seeds on an opinion network whose effects propagate through coupled opinion and epidemic dynamics. The model places the same individuals on a square-lattice disease layer and a Barabási–Albert opinion layer, with opinion states $\tau$5 influencing vaccination decisions and disease outcomes. Seed selection is static at $\tau$6, but its consequences are dynamic because opinions and infections coevolve thereafter. Simulations show that network-based targeting with pro-vaccine supporters as initial seeds increases vaccine adoption and reduces epidemic size, and that choosing individuals with central positions in the opinion network is much more effective than community-based or profession-based targeting; hubs perform best, and betweenness-based targeting is also strong (Fügenschuh et al., 2023).

A complementary graph-theoretic literature studies competitive opinion targeting under DeGroot consensus with stubborn agents. Regular-node steady states satisfy

$\tau$7

and the plus-agent seeks a target set maximizing the average steady-state opinion

$\tau$8

The set function is submodular, so greedy enjoys the usual $\tau$9 guarantee, but the paper derives sharper structure-dependent rules. On complete graphs it provides an analytical solution showing when it is better to block the opponent by targeting the same nodes. On line graphs it solves the single-targeting problem in closed form. On trees it proves that the objective is unimodal along any branch and proposes the Tree Graph Single Targeting Algorithm, which returns the exact STP solution on trees and works well as a heuristic on sparse or tree-like graphs (Bini et al., 2021).

These network papers are narrower than the spacecraft or marketing definitions, but they show an important variant: target selection may itself be one-shot while the targeted effect is fundamentally dynamic because it is mediated by diffusion, consensus, or contagion on a graph.

6. Limitations, failure modes, and emerging directions

Across domains, dynamic targeting is usually constrained by a narrow feasible envelope between responsiveness and instability. On CogniSAT-6, the window between lookahead imaging and nadir can be only 60–90 s, image readout was not designed for seconds-scale DT, and practical mitigations include decimation, reduced spectral subsets, and speculative slews (Chien et al., 22 Aug 2025). In laser tuning, continuous operation depends on maintaining the phase–strength relation closely enough to remain on the maximum gain mode; incorrect ratios yield mode hops or spectral broadening, and the paper identifies on-chip feedback control as a natural next step (Mey et al., 3 Jun 2025). In privacy-preserving targeting, the effectiveness of strategic querying deteriorates when the CATE surface is very jagged, the query budget is extremely small, or the DP noise scale is very high, even though the method remains strong in realistic intermediate-budget regimes (Shchetkina, 7 Jul 2025).

A second recurrent limitation is that several systems marketed as dynamic are only partially adaptive. The multiplex vaccine study fixes seeds once at initialization and explicitly notes that static seeding on static networks is a limitation, motivating future adaptive targeting on evolving networks (Fügenschuh et al., 2023). EGL updates its entity graph weekly and user preferences daily, but the authors still identify distribution shift, hierarchical graph structure, and richer temporal modeling as open directions (Yang et al., 2023). The constrained marketing LP is powerful for repeated re-solving, yet the formulation itself is per-period rather than explicitly multi-period, and the paper points toward integration with online learning or richer sequential formulations (Lu et al., 2023).

The broader trajectory suggested by these works is toward tighter coupling of targeting, feedback, and coordination. The spacecraft work points to cross-platform tasking through inter-satellite or low-latency links (Chien et al., 22 Aug 2025). Privacy-preserving targeting points to Bayesian querying combined with contextual or multitask bandits (Shchetkina, 7 Jul 2025). Target-aligned data selection suggests continual reuse and recalibration of compact target-specific warmups across new candidate pools (Yang et al., 10 May 2026). Inference from this convergence is straightforward: dynamic targeting is increasingly less about a static ranking rule and more about maintaining a target-conditioned control loop under timing, privacy, compute, and fairness constraints.