Unpredictable Control Methods

• Unpredictable control is a family of paradigms that deliberately embed randomness—through algorithmic, hardware-native, or dynamical means—to obscure system trajectories from observers.
• It employs methods like temporal-logic synthesis, entropy maximization, and stochastic actuation to balance security requirements with control performance.
• Applications range from secure synthesis and distributed robotics to quantum LiDAR and chaos management, illustrating trade-offs between control stability and unpredictable behavior.

Searching arXiv for recent and foundational papers on unpredictable control to ground the article. Unpredictable control denotes a class of control problems in which unpredictability is itself a design objective, a security constraint, an emergent dynamical property, or a disturbance characteristic that must be managed rather than eliminated. Across the recent literature, the term includes controllers that prevent an observer from determining exactly when a task will be completed (Chen et al., 2022), policies that maximize trajectory entropy under reward constraints in partially observable models (Hibbard et al., 2019), stochastic actuation schemes that deliberately enlarge an adversary’s prediction error (Qu et al., 2022), nonlinear and chaotic systems whose forcing or dissipation changes whether basin structure is Wada and hence difficult to predict (Coccolo et al., 2023), and hardware-native mechanisms in which the control variable is selected by intrinsic quantum randomness rather than a deterministic steering signal (Kim et al., 12 Nov 2025). The resulting field is therefore not organized around a single metric; it is organized around the relation between controllability, observability, inference, and the deliberate or intrinsic production of uncertainty.

1. Conceptual scope and principal meanings

In the literature represented here, unpredictable control appears in at least four technically distinct senses. First, unpredictability can be a security property: the controller must achieve a task while denying an observer exact knowledge of future completion time or future trajectory. This is the setting of temporal-logic synthesis under passive observation and entropy-maximizing planning under partial observability (Chen et al., 2022). Second, unpredictability can be a designed output property: random inputs are injected so that future outputs or states become irregular and harder to infer, while nominal control objectives remain satisfied (Qu et al., 20 Aug 2025). Third, unpredictability can be an emergent feature of nonlinear dynamics: Wada basin boundaries, chaotic transients, or noise-driven switching can make future evolution highly sensitive to perturbations even when the governing equations are known (Coccolo et al., 2023). Fourth, unpredictability can be hardware-native rather than algorithmically imposed, as in beam steering by quantum wavelength randomness, where no classical drive signal sets the direction in advance (Kim et al., 12 Nov 2025).

A further distinction concerns whether unpredictability is regarded as desirable or undesirable. In chaotic scattering, the control problem may be to mitigate unpredictability by dissipation that recomposes basin boundaries (Coccolo et al., 2023). In secure synthesis and sensing, the goal is instead to guarantee unpredictability against malicious inference (Chen et al., 2022). In human motor control, unpredictability may reflect an intrinsic stochastic activation mechanism rather than an engineered design choice: the onset of corrective action is modeled as a stochastic escape from an idle manifold rather than threshold crossing (Zgonnikov et al., 2014).

This distribution of meanings suggests that unpredictable control is best understood as a family of inference-aware control paradigms. The common question is not whether the system is random, but which variables are meant to be predictable, by whom, over what horizon, and under which performance constraints.

2. Formal synthesis under observation and adversarial inference

A particularly explicit formulation is given in controller synthesis for temporal-logic tasks under an eavesdropper. The system is a non-deterministic transition system

$G = (X, x_0, U, \to, AP, L),$

and the task is expressed by an scLTL formula. A controller is required to keep the closed-loop system live, ensure eventual finite-time satisfaction of the task, and preserve $K$-step unpredictability. A finite path is $K$-step unpredictable if the intruder cannot determine for sure that the system will exactly accomplish the task in $K$ steps ahead; equivalently, it suffices to check that for every closed-loop path there exists an observation-indistinguishable path of length $K$ whose last state is not in the accepting set of the product system (Chen et al., 2022). The key construction augments states with a prediction vector $h\in\{0,1\}^{K+1}$, builds secure belief states, and synthesizes a controller through a bipartite transition system and a reachability game. The resulting algorithm is sound and complete (Chen et al., 2022).

Reactive Control Improvisation generalizes this perspective to finite-horizon reactive synthesis with an explicit randomness requirement. The controller must satisfy a hard constraint $H\subseteq\Sigma^n$, a soft constraint $S\subseteq\Sigma^n$, and a bound that no single play be generated with probability exceeding $\rho$. With $I=H$ and $K$0, realizability is characterized by the width conditions

$K$1

where $K$2 measures how many improvisations can be guaranteed against an adversary (Fremont et al., 2018). This formalizes unpredictability as bounded concentration of path probabilities rather than as state noise.

In partially observable Markov decision processes, unpredictability is often quantified by trajectory entropy. The objective is

$K$3

For finite-state controllers, the problem is recast as parameter synthesis for a parametric Markov chain, and a nonlinear optimization solved by a penalty-based convex–concave procedure yields locally optimal controllers of fixed memory size (Hibbard et al., 2019). A notable theorem states that a decision-maker with perfect observations can randomize its paths at least as well as a decision-maker with partial observations:

$K$4

This rules out the common intuition that partial observability itself creates stronger unpredictability; it may hinder control randomization as much as it hides it (Hibbard et al., 2019).

3. Randomization and noise as control resources

For linear mobile agents, unpredictability can be engineered by decomposing the input as

$K$5

where $K$6 is nominal and $K$7 is zero-mean random perturbation. Against a Kalman-filter-style attacker, the one-step prediction error is

$K$8

and the design goal is to maximize

$K$9

while retaining quadratic control performance. Under the variance prescription used in the paper, the worst-case one-step prediction error is achieved when each perturbation component is uniformly distributed,

$K$0

and the resulting multi-period problem is convex and solved through dynamic programming, including an input-constrained extension (Qu et al., 2022). In the scalar example, the average one-step prediction error rises from $K$1 with no perturbation to $K$2 when $K$3 (Qu et al., 2022).

A related linear-systems formulation introduces two explicit unpredictability metrics:

$K$4

where $K$5. Under the variance metric, the worst-case design problem has the analytic solution $K$6: the optimal symmetric distribution is any law with covariance equal to the upper covariance bound (Qu et al., 20 Aug 2025). Under the confidence-probability metric, the original optimization is non-convex, but symmetry, monotonicity, and discretization reduce it to a linear program. In a second-order integrator example with identical covariance $K$7, the LP-computed optimal density yields smaller $K$8 than uniform, Gaussian, or Laplace densities: $K$9, $K$0, and $K$1 at $K$2, $K$3, and $K$4, respectively (Qu et al., 20 Aug 2025).

Noise can also improve stabilization rather than merely obscure inference. In noisy prediction-based control,

$K$5

bounded i.i.d. perturbations of the control coefficient can globally stabilize the unique equilibrium $K$6 in cases where the deterministic mean control $K$7 does not ensure global stability and may admit a stable two-cycle (Braverman et al., 2023). For unimodal maps with negative Schwarzian derivative, the paper extends Singer-type “local implies global” results to the stochastic setting (Braverman et al., 2023).

Quantum control offers a distinct use of randomness. In random dynamical decoupling over quantum open systems, pulse strength, duration, and timing are perturbed by

$K$8

for $K$9. Using an exact quantum-state-diffusion treatment, Jun Jing, C. Allen Bishop, and Lian-Ao Wu show that out-of-order random pulses work as well as regular pulses for dynamical decoupling and dissipation suppression in a large parameter error region, provided the average pulse rate is sufficiently high and the environment is non-Markovian (Jing et al., 2014). The reported effective regime includes $K$0, $K$1, and duty cycle $K$2 (Jing et al., 2014).

4. Nonlinear dynamics, chaos, and control of unpredictability

In open nonlinear Hamiltonian systems, unpredictability is often tied to the geometry of exit basins. For the randomly driven Hénon–Heiles system, the conservative Hamiltonian

$K$3

is supplemented by linear damping and periodic forcing with random phase. The full equations include

$K$4

The control question is whether forcing amplitude $K$5, damping $K$6, and energy $K$7 place the system in a Wada or non-Wada regime (Coccolo et al., 2023). At fixed $K$8, a numerically determined boundary $K$9 separates Wada from non-Wada basins; example points are $h\in\{0,1\}^{K+1}$0, $h\in\{0,1\}^{K+1}$1, and $h\in\{0,1\}^{K+1}$2, $h\in\{0,1\}^{K+1}$3 (Coccolo et al., 2023). With zero forcing, the minimum dissipation for Wada basins is fitted by

$h\in\{0,1\}^{K+1}$4

valid for $h\in\{0,1\}^{K+1}$5, with $h\in\{0,1\}^{K+1}$6 and $h\in\{0,1\}^{K+1}$7 (Coccolo et al., 2023). The mean escape time decreases from $h\in\{0,1\}^{K+1}$8 in a non-Wada regime to $h\in\{0,1\}^{K+1}$9 in a Wada regime, indicating that stronger forcing both accelerates escape and increases mixing (Coccolo et al., 2023).

A complementary problem is not to avoid escape from a transient chaotic region but to force it in a predictable manner. Alfaro, Capeáns, and Sanjuán consider

$H\subseteq\Sigma^n$0

with bounded disturbance and chaotic transient dynamics in a compact region $H\subseteq\Sigma^n$1. They define an escape function $H\subseteq\Sigma^n$2 and corresponding escape sets $H\subseteq\Sigma^n$3, so that trajectories can be driven out of $H\subseteq\Sigma^n$4 in at most or exactly $H\subseteq\Sigma^n$5 steps depending on the recursion used (Alfaro et al., 2021). In a logistic-map example with $H\subseteq\Sigma^n$6, $H\subseteq\Sigma^n$7, $H\subseteq\Sigma^n$8, and $H\subseteq\Sigma^n$9, uncontrolled escape times vary fractally from $S\subseteq\Sigma^n$0 to more than $S\subseteq\Sigma^n$1 steps, whereas controlled trajectories in $S\subseteq\Sigma^n$2 exit in $S\subseteq\Sigma^n$3 or exactly $S\subseteq\Sigma^n$4 steps (Alfaro et al., 2021).

Menon et al. describe another chaos-based strategy, chaotic-trap control, for driven nonlinear maps. A simple threshold test switches the drive between $S\subseteq\Sigma^n$5 and $S\subseteq\Sigma^n$6,

$S\subseteq\Sigma^n$7

creating a trap interval $S\subseteq\Sigma^n$8 that can be made arbitrarily small as $S\subseteq\Sigma^n$9 shrinks (Menon et al., 2015). Inside the trap, the input sequence becomes chaotic with Bernoulli-like statistics, sub-sequence probabilities $\rho$0, and entropy approaching $\rho$1 (Menon et al., 2015). Here unpredictability of the switching signal is not a defect but the mechanism that keeps the state confined.

5. Learning, sensing, and distributed robotics

In unsupervised reinforcement learning, unpredictability interacts with controllability. Adversarial Surprise defines “surprise” by $\rho$2 and sets up a zero-sum game between an Explore policy and a Control policy that alternate control of the same body. The Control agent maximizes

$\rho$3

while the Explore agent receives $\rho$4 (Fickinger et al., 2021). In stochastic Block MDPs with disjoint observation support and a $\rho$5-cover-by-dark-rooms assumption, the induced process provably $\rho$6-covers the latent state space, thereby avoiding both the “noisy-TV” pathology of pure exploration and the “dark-room” pathology of pure surprise minimization (Fickinger et al., 2021).

A related learning-enabled but safety-centered direction is cognitive-flexible control with latent model reorganization. CF–DeepSSSM maintains a stochastic latent belief and bounds online representation change by a Cognitive Flexibility Index,

$\rho$7

then embeds the adapted model in Bayesian MPC with chance constraints (Nuchkrua et al., 31 Jan 2026). The paper establishes bounded posterior drift, recursive feasibility, closed-loop ISS, and a safety-preservation corollary. In the reported dynamics-shift scenario over $\rho$8 steps, Safety Rate is $\rho$9, cumulative tracking cost is approximately $I=H$0 versus $I=H$1 for nominal MPC and $I=H$2 for robust MPC, and mean CFI is approximately $I=H$3 (Nuchkrua et al., 31 Jan 2026). This line is not primarily about hiding future behavior from an observer; it addresses safe control under abrupt nonstationarity while preserving explicit guarantees.

Distributed robotics emphasizes robustness to unpredictable targets and environments. ISEE.U performs online active target localization from noisy ranges without assuming target dynamics. Each agent greedily selects motion by D-optimality using a local Fisher-information surrogate, and the distributed estimate is unbiased at every consensus iteration (Vasques et al., 2022). In experiments with an unpredictable spiral target, the benchmark particle-filter method of Meyer et al. deteriorates after approximately $I=H$4 steps, whereas ISEE.U’s MAE continues decreasing to near zero; the method also uses $I=H$5 less computation time (Vasques et al., 2022). For connectivity recovery after unpredictable obstacle insertions, a distributed prediction-based replanning method restores a unique relay chain from base to goal whenever a solution exists. Over $I=H$6 random trials, $I=H$7 solvable, the reported mean mission times are $I=H$8 s for the Full-information baseline, $I=H$9 s for Search, and $K$00 s for the prediction-based method; corresponding distances are approximately $K$01 m, $K$02 m, and $K$03 m (Marchukov et al., 14 Mar 2025).

Quantum LiDAR provides a particularly sharp instance of inherently unpredictable actuation. Photon pairs are generated by spontaneous four-wave mixing; the probe photon is diffracted at angle $K$04 according to

$K$05

while the herald time encodes wavelength through dispersion,

$K$06

Because each pair’s wavelength is selected by vacuum fluctuations, no controller sets the beam direction in advance; the direction is known only after herald detection (Kim et al., 12 Nov 2025). The reported system detects $K$07 targets in one acquisition, with total acquisition time $K$08 s, range resolution $K$09 cm, angular resolution $K$10, and signal-to-noise enhancement up to $K$11 relative to classical LiDAR under the stated comparison (Kim et al., 12 Nov 2025).

6. Guarantees, trade-offs, and recurrent misconceptions

A recurrent misconception is that unpredictability is synonymous with unconstrained randomness. The formal synthesis literature shows the opposite. In temporal-logic synthesis, unpredictability is enforced simultaneously with liveness and eventual task completion (Chen et al., 2022). In POMDP planning, entropy maximization is constrained by expected reward, and partial observation does not dominate full observation in achievable entropy (Hibbard et al., 2019). In Bayesian MPC with latent reorganization, representation change is explicitly bounded by the CFI constraint and coupled to recursive-feasibility and safety guarantees (Nuchkrua et al., 31 Jan 2026).

A second misconception is that noise necessarily degrades control. Several results are explicitly contrary. Bounded stochastic perturbations of a PBC gain can induce a stability switch from two-cycle or chaotic behavior to almost-sure convergence (Braverman et al., 2023). Random pulse sequences can remain robust for dynamical decoupling under substantial amplitude, width, and timing fluctuations (Jing et al., 2014). At the same time, these papers do not imply that more randomness is always preferable. In linear secure control, the same covariance that increases unpredictability also raises LQR or cooperative-control cost through terms such as

$K$12

making the privacy–performance trade-off explicit (Qu et al., 20 Aug 2025).

A third misconception is that unpredictability is always a design objective. In the randomly driven Hénon–Heiles system, external random-phase forcing is identified as the main source of unpredictability, while dissipation can restore smooth, compact, non-Wada basins (Coccolo et al., 2023). In human intermittent control, the central issue is not deliberate concealment but whether corrective actions are threshold-driven or intrinsically stochastic. Experiments on virtual stick balancing show that action-point distributions decay nearly exponentially and contradict threshold models with approximately Gaussian action-point clustering; the noise-driven model reproduces the observed exponential law (Zgonnikov et al., 2014). In conceptual climate dynamics, generalised synchronisation to astronomical forcing can coexist with weak structural stability, so that small parameter changes or random fluctuations may shift the sequence of ice ages; a high Rayleigh number on eccentricity is explicitly stated to be no guarantee of reliable synchronisation (Crucifix, 2013).

Taken together, these works frame unpredictable control as a technically heterogeneous but conceptually coherent domain. The unifying theme is not randomness per se, but the controlled management of predictability: when it should be reduced, when it should be restored, how it is quantified, and what formal guarantees can be maintained while doing so.