Dynamic Targeting (DT): Adaptive Control Methods

• Dynamic Targeting (DT) is an adaptive framework that continuously updates objectives based on real-time feedback, residual error, or lookahead data.
• DT techniques are applied across diverse fields such as semiconductor laser tuning, semantic genetic programming, adaptive drug dosing, satellite observation, and blockchain consensus.
• Practical DT implementations demonstrate enhanced system stability, improved predictive accuracy, and optimized resource allocation through tailored control strategies.

Dynamic Targeting (DT) is a term used in several technical literatures for adaptive schemes in which a target, subtarget, or observation objective is updated during operation rather than fixed in advance. The label appears in semiconductor-laser frequency control, semantic genetic programming, adaptive drug dosing, satellite observation planning, privacy-preserving personalization, proof-of-work consensus, dynamic treatment allocation in Markovian systems, and oligopoly advertising (Mey et al., 3 Jun 2025, Ruberto et al., 2020, Engelhardt, 2019, Kangaslahti et al., 5 Mar 2026, Shchetkina, 7 Jul 2025, Bissias, 2020, Hu et al., 30 Jun 2025, Fletcher et al., 2015). The methods are not algorithmically identical. They range from coordinated optical feedback and residual-error recursion to continuous-action reinforcement learning and time-varying mining thresholds. This suggests that DT is best understood as a cross-domain label for adaptive targeting architectures rather than as a single formalism.

1. Semiconductor-laser dynamic targeting

In photonics, DT denotes a method for agile, fast, and continuous wavelength tuning of semiconductor lasers by coordinated control of optical-feedback rate $\kappa$ and phase $\phi$. The formal model is the normalized Lang–Kobayashi system for the complex field $E(t)$ and carrier density $N(t)$,

$$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$

With feedback turned on, steady-state external cavity modes appear. Among them, the Maximum-Gain Mode (MGM) is the unique solution whose frequency shift $\Delta\omega$ increases linearly with $\kappa$ and does not undergo hop-offs as $\kappa$ is raised. Under the classical condition $\phi=\tau\alpha\kappa$, linearization yields

$$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$

More generally, stability analysis gives

$\phi$0

under which the MGM is the dominant attractor and side-modes remain suppressed. DT consists of continuously enforcing that ratio as $\phi$1 is ramped, thereby locking the laser on a single mode and suppressing discrete mode hops (Mey et al., 3 Jun 2025).

The experimental proof-of-principle uses a DFB edge-emitter laser at $\phi$2 nm at 60 mA and 25 °C in a free-space setup. A neutral-density-filter wheel provides coarse feedback-strength setting; a quarter-wave plate and linear polarizer on a motorized rotation stage provide fine dynamic control of $\phi$3 via a $\phi$4 transmission law; and a high-reflectivity mirror on a Newport XMS50 linear stage sets $\phi$5 by sub-wavelength position changes. In real time, $\phi$6 is varied by rotating the QWP–LP assembly, while $\phi$7 is updated via synchronized mirror motions to keep $\phi$8. In the free-space implementation this synchronization is obtained iteratively by calibration; the paper notes that an integrated loop could instead use a fast piezo-actuator or on-chip phase shifter driven by the same control voltage as $\phi$9.

Experimentally, the center wavelength shifts continuously by $E(t)$0 GHz, approximately 17 pm at 1550 nm, without any observable mode hops. Sweeping $E(t)$1 alone or $E(t)$2 alone instead produces periodic hops or chaotic multimode emission. Bidirectional tuning shows no measurable hysteresis: the wavelength-versus-mirror-position curves overlap within the $E(t)$3 MHz experimental resolution, and over more than 20 cycles the start-and-end wavelengths coincide to within $E(t)$4 MHz. Residual wavelength “wiggling” of $E(t)$5 MHz is attributed to mechanical micro-vibrations. Simulations with $E(t)$6, $E(t)$7, $E(t)$8, triangular $E(t)$9, and $N(t)$0 predict a maximum normalized shift of about $N(t)$1, corresponding to approximately 120 GHz in real units using a 5 ps photon lifetime, and a scan speed of about $N(t)$2. The same study reports that shorter cavities are generally more favorable for fastest tracking and identifies coherent telecom, LIDAR/sensing, and high-resolution spectroscopy as likely applications.

2. Residual-driven dynamic targets in semantic genetic programming

In semantic genetic programming, Dynamic Target refers to redefining the regression target after each run to be the unexplained residual of the current best model. If $N(t)$3 is the target vector at external iteration $N(t)$4, initialized as $N(t)$5, and $N(t)$6 is the best model found in run $N(t)$7, then

$N(t)$8

Each new run therefore focuses on whatever the previous run failed to explain. The method combines DT with linear scaling. For a candidate program $N(t)$9, the scaled form is

$$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$0

where $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$1 and $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$2 are computed by least squares against the current target, and the fitness is

$$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$3

The final model is the sum of the retained linearly scaled partial models,

$$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$4

(Ruberto et al., 2020).

The workflow consists of chaining $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$5 short GP runs of length $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$6, with a fresh population at each run. The implementation described in the paper uses population size 1,000 and a total budget of 1,000 generations, instantiated as $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$7 runs and $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$8 generations. It is mutation-only: there is no classic crossover, and each generation uses tournament selection and 1-elitism. A 10% validation holdout is used for selecting which partial models to keep in the final chain.

On eight benchmark data sets—Airfoil, Concrete, EN–Cooling, EN–Heating, Housing, Tower, Yacht, and uBall5D—SGP-DT is compared against lasso regression, $$\dot{E}(t) =\frac{1}{2}\bigl[1 - i\alpha\bigr]\,N(t)\,E(t) +\kappa\,E(t - \tau)\,e^{\,i\phi}, \qquad \dot{N}(t) =\frac{P - N(t) - \bigl[1 + 2\,N(t)\bigr]|E(t)|^2}{T}.$$9-lexicase selection, and two ablations. The reported median test-RMSE reductions are 51.47% versus lasso, 23.19% versus lexicase, 5.39% versus DT-EM, and 3.31% versus DT-NM. In computational effort, SGP-DT requires on average $\Delta\omega$0 fewer node-evaluations than lexicase, approximately the same effort as DT-EM ($\Delta\omega$1), and about $\Delta\omega$2 more than DT-NM. The paper also reports limitations: some overfitting on very small data sets such as Yacht and Housing, final trees that can be hundreds of nodes, and a fixed number of external runs that may not be optimal for every problem. Within this literature, DT is therefore a residual-refocusing mechanism rather than a feedback controller.

3. State-feedback targeting in medicine and dynamic treatment systems

In adaptive drug dosing, DT has been used for automated precision control of stochastic and heterogeneous cell proliferation under emergent resistance. The CelluDose formulation models phenotype densities $\Delta\omega$3 under drug concentrations $\Delta\omega$4 with a Hill-type dose-response birth rate, logistic crowding, demographic stochasticity, and rare discrete mutation events. Control is posed at decision epochs every $\Delta\omega$5 h, with state

$\Delta\omega$6

continuous action

$\Delta\omega$7

and an objective that combines elimination of all cell types by a finite horizon with minimization of cumulative drug usage. The learning algorithm is Deep Deterministic Policy Gradient (DDPG), with a continuous-action actor and critic, replay buffer $\Delta\omega$8, batch size 128, actor learning rate $\Delta\omega$9, critic learning rate $\kappa$0, discount $\kappa$1, and OU exploration noise. Dynamic targeting occurs at feedback time: when only susceptible cells remain, the learned policy applies a low baseline dose; when a resistant subpopulation emerges, dosing increases proportionally to the size and resistance level of that subpopulation, and in combination therapy the controller may switch from a first-line to a last-line drug and then revert once the resistant clone is cleared (Engelhardt, 2019).

The reported empirical behavior is unusually strong. In the single-drug setting, 1,000 new episodes with at least one mutation all end in cure, and the controller remains successful even when the mutation probability is raised to $\kappa$2, which is 100$\kappa$3–1,000$\kappa$4 above the training value $\kappa$5. In the dual-drug setting, success remains 100% up to $\kappa$6 and then falls to about 47% at $\kappa$7. The same study reports robustness to variations in mutation rate, initial population size from $\kappa$8 to $\kappa$9 cells/mL, and stochastic noise magnitudes, with convergence after about 50,000 episodes in the single-drug case and about 130,000 episodes in the dual-drug case.

A distinct treatment-allocation literature uses DT in Markovian systems with shared capacity constraints. Here units arrive sequentially with covariates $\kappa$0, system state $\kappa$1, binary treatment $\kappa$2, and outcome $\kappa$3, while the state evolves according to a Markov kernel $\kappa$4. The objective is long-run average reward

$\kappa$5

The key result is that the optimal policy compares a CATE-like direct effect,

$\kappa$6

to a state-specific threshold $\kappa$7 reflecting the expected cumulative indirect effect on later arrivals. The optimal rule is

$\kappa$8

with

$\kappa$9

Estimation proceeds by combining a standard CATE estimator with off-policy evaluation over a grid of state-wise quantile levels, and the main theorem gives regret rates of $\phi=\tau\alpha\kappa$0 under $\phi=\tau\alpha\kappa$1 CATE error and $\phi=\tau\alpha\kappa$2 under sup-norm CATE error. In queueing experiments, system-aware DT yields up to 15–20% lift in steady-state rewards relative to direct CATE targeting, and converges to the true MDP optimum as sample size grows (Hu et al., 30 Jun 2025).

These two literatures use the same label for different control logics. In CelluDose, DT means escalation and de-escalation of dosage in response to measured subpopulation composition. In Markovian treatment allocation, DT means thresholding individual-level benefit by a state-dependent congestion cost. The shared element is feedback from system state into the targeting decision.

4. Earth-observation planning and spacecraft autonomy

In satellite observation, DT is a mission concept in which a lookahead sensor gathers information about the upcoming environment and an onboard planner uses that information to direct a higher-value primary sensor. The formal planning problem uses a state

$\phi=\tau\alpha\kappa$3

where $\phi=\tau\alpha\kappa$4 indexes the current footprint, $\phi=\tau\alpha\kappa$5 is the number of observations taken so far, and $\phi=\tau\alpha\kappa$6 is cycle count, together with an action that specifies a reachable new aimpoint and a Boolean decision to observe. The objective is

$\phi=\tau\alpha\kappa$7

Because the naive search space scales as $\phi=\tau\alpha\kappa$8 and $\phi=\tau\alpha\kappa$9 can be on the order of 450 to 525 cycles, the 2026 hierarchical-planning work uses supplemental geostationary data to build a long-term “blueprint” and then refines it with onboard lookahead. Geostationary streams provide lookahead up to 35 minutes ahead of time, compared with about 1 minute from an onboard lookahead sensor. The long-term stage partitions the trajectory into windows, computes a utility proxy $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$0 for each window, and allocates the observation budget proportionally in $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$1 time or $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$2 if $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$3 is the average number of pixels per window. The short-term stage uses onboard data with beam search of depth $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$4 and beam width $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$5 (Kangaslahti et al., 5 Mar 2026).

The benchmark scenarios are Cloud Avoidance (CA), Cloud Avoidance + Population Density (CAPD), Cloud Avoidance + Random Targets (CART), and Storm Hunting (SH). Performance is measured as aggregate utility relative to an omniscient upper bound that sums the top 100 reachable-pixel utilities ignoring slewing. The Geostationary Satellite Data-Informed hierarchical planner achieves up to 99.9% of the upper bound in CA, 72.8% in CAPD, 63.3% in CART, and 67.2% in SH, and it outperforms the best onboard-only hierarchical variant by up to 41%. Relative to the best Greedy Historical baseline, the same planner approximately doubles average utility, with an average factor of about 2.44 across scenarios. The gains are largest when high-utility events are dynamic and sparsely distributed over the overflight, such as storms or population-weighted clear-sky opportunities.

A flight-oriented DT architecture is described for the CogniSAT-6 spacecraft. There, DT is defined as an autonomy concept in which onboard sensors acquire lookahead data, analyze it rapidly, and use the result to drive immediate retargeting of a higher-performance primary sensor. CogniSAT-6 does not carry a dedicated lookahead imager; instead, the primary multispectral camera is time-shared. The operational loop has five steps: acquire the lookahead image, transfer it to the edge processor, detect an event or cloud-free corridor, generate retargeting commands, slew by about 40–50° along track, and acquire the follow-up nadir image. The system uses an Intel Movidius Myriad X Vision Processing Unit with 1 GB LPDDR and 4 GB flash under real-time Linux, with a latency budget

$$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$6

decomposed into transfer, preprocessing, analysis, command generation, slew, nadir acquisition, and downlink. Implemented onboard pipelines include radiometric stretching, cloud detection by brightness threshold and morphological opening, thermal-anomaly detection by Spectral Angle Mapper, and a lightweight CNN for convective storm classification. In-orbit test results reported so far include 50° slews in about 35 s with $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$7 pointing jitter, image transfer plus analysis plus command issue in about 18 s, cloud detection precision 0.96 and recall 0.94, thermal-anomaly precision 0.90 and recall 0.88, average analysis latency 8.5 s, retargeting latency 20 s, and a 30% increase in cloud-free pixels relative to a non-DT baseline. Additional DT metrics include $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$8 in usable science frames per orbit for cloud avoidance, $$\Delta\omega=\omega-\omega_0=-\,\alpha\,\kappa, \qquad \Delta f=\frac{\Delta\omega}{2\pi}\approx -\,\frac{\alpha}{2\pi}\,\kappa.$$9 and $\phi$00 for storm hunting, and plume-detection latency below 90 s with a goal below 60 s (Chien et al., 22 Aug 2025).

5. Privacy, blockchain, and market-allocation variants

A privacy-preserving targeting literature uses DT for dynamic data exploration under third-party privacy constraints. Instead of raw individual data, the platform exposes only a limited number of aggregate queries on hyperrectangles of the covariate space and adds Gaussian differential-privacy noise with $\phi$01. The latent treatment-effect surface is modeled as a zero-mean Gaussian process, but the observations are noisy integrals over regions rather than point evaluations. To select the next query, the paper introduces a targeting-aware acquisition function. The conceptual version is

$\phi$02

while the algorithmic version uses

$\phi$03

The method is described as strategic querying and is reported to outperform uniform querying when the query budget is moderate, the DP noise scale is non-negligible, and the latent treatment-effect surface has intermediate smoothness. On the Criteo AI Labs uplift data with 14M users and cost $\phi$04, strategic querying achieves 97–101% of the non-privacy-preserving Causal Forest benchmark across four privacy settings, whereas uniform querying ranges from 97% down to 33% of that benchmark (Shchetkina, 7 Jul 2025).

In proof-of-work consensus, Dynamic PoW Targeting refers to any mechanism in which the mining threshold is adjusted continuously as time elapses since the last block. Real-Time Block Rate Targeting (RTT) uses a subtarget

$\phi$05

so that inter-arrival times are Weibull distributed with expected block interval $\phi$06. For the common choice $\phi$07, the variance ratio is

$\phi$08

corresponding to roughly a 73% reduction in block-time variance relative to conventional PoW. RTT, however, is vulnerable to future-mining: miners can precommit to a future subtarget and obtain a per-hash success probability greater than their fair hash-rate share. Radium modifies RTT by scaling rewards as

$\phi$09

which makes expected reward per hash constant throughout the mining interval and restores incentive compatibility. A one-block feedback rule updates difficulty every block in the transformed exponential domain. Simulations over 30 blocks with retarget every block and $\phi$10 show 5 s median deviation around $\phi$11 s and a 95th-percentile delay within 18% of the ideal exponential. Large-scale simulations with a 3 s propagation window find orphan rates of 0.22% for Bitcoin and 0.36% for Radium, while doublespend success for a 30% attacker and a 6-block embargo is approximately 0.15 for both protocols (Bissias, 2020).

In oligopoly advertising, DT denotes targeted advertising effort allocation between two categories of prospective customers: the uncommitted market and customers of rival firms. If $\phi$12 is firm $\phi$13’s market share and $\phi$14 is market potential, then the controls are $\phi$15 for the uncommitted market and $\phi$16 for competitors’ customers. The state dynamics are

$\phi$17

Each firm maximizes discounted profit with quadratic effort costs. In the duopoly case, the equilibrium controls are

$\phi$18

Numerical examples show a characteristic time path: early in the game, when $\phi$19 is large, effort toward the uncommitted market dominates; as $\phi$20 decays, firms shift toward offensive effort against rivals’ customers. Closed-loop DT adapts to unexpected swings in market share, whereas open-loop DT can perform poorly under disturbance. For a firm with a fixed instantaneous budget $\phi$21, Theorem 5.4 gives a closed-form optimal split that depends only on the firm’s own $\phi$22 ratio and the current competitor share $\phi$23 (Fletcher et al., 2015).

6. Cross-domain interpretation and recurrent distinctions

A common misconception is that Dynamic Targeting names a single transferable algorithm. The cited literature instead uses the term for several distinct operations: maintaining a semiconductor laser on the Maximum-Gain Mode by coordinating $\phi$24 and $\phi$25, replacing a regression target by residual error in semantic GP, and planning future observations from lookahead data in satellite autonomy (Mey et al., 3 Jun 2025, Ruberto et al., 2020, Kangaslahti et al., 5 Mar 2026). Another misconception is that “targeting” necessarily refers to individual treatment assignment. In these literatures it can refer to a wavelength, a residual vector, a resistant subpopulation, a satellite aimpoint, a covariate hyperrectangle under differential privacy, a mining threshold, or a segment of the advertising market (Engelhardt, 2019, Shchetkina, 7 Jul 2025, Bissias, 2020, Fletcher et al., 2015).

Despite this heterogeneity, several recurrent structures are visible. Many DT systems augment the instantaneous state with a compressed statistic that drives control: $\phi$26 in CelluDose, state-specific thresholds $\phi$27 in Markovian treatment allocation, utility windows in hierarchical satellite planning, and dynamic subtargets $\phi$28 in PoW (Engelhardt, 2019, Hu et al., 30 Jun 2025, Kangaslahti et al., 5 Mar 2026, Bissias, 2020). Others rely on explicit residualization or decomposition, such as $\phi$29 in SGP-DT and long-term blueprint plus short-term refinement in satellite planning (Ruberto et al., 2020, Kangaslahti et al., 5 Mar 2026). This suggests a broad design pattern: DT methods typically replace a fixed global objective with an adaptive local objective whose update rule is tied to residual error, measured state, or lookahead information.

The principal difference across domains is therefore not whether the target changes, but what object is being updated and under what constraints. In photonics the update rule must preserve mode stability; in genetic programming it must preserve semantic progress; in medicine it must handle stochastic mutation and toxicity; in satellite systems it must satisfy real-time compute and slewing limits; in privacy-preserving personalization it must use only noisy aggregates; in proof-of-work it must remain incentive-compatible; and in advertising it must balance offensive and market-expansion effort over time. The term “Dynamic Targeting” is consequently best read as a domain-specific technical label whose meaning depends on the surrounding mathematical model, control interface, and operational constraints.