Human-in-the-Loop Optimisation (HILO)

Updated 9 October 2025

Human-in-the-Loop Optimisation is a hybrid paradigm that integrates human expertise with computational algorithms to address uncertainty and complex constraints.
It employs methodologies such as Bayesian optimization, evolutionary strategies, and real-time physiological feedback to guide and refine solution search.
Practical applications in robotics, assistive devices, and personalized interfaces highlight its effectiveness while addressing challenges like human variability and convergence issues.

Human-in-the-Loop Optimisation (HILO) is a hybrid paradigm in optimization and system control that integrates real-time human expertise and judgment into computational search, modeling, and control algorithms. This approach aims to solve problems where fully automated optimization is infeasible due to uncertain objectives, complex constraints, incomplete formalization, or high individual variability. HILO systems employ iterative loops in which human insight (e.g., selection, feedback, correction, or supervision) guides or constrains the optimization process, leveraging the strengths of both parties to improve solution relevance, safety, interpretability, and individual personalization.

1. Foundational Principles of HILO

Human-in-the-Loop Optimisation is based on the collaboration between autonomous agents (algorithms, robots, or controllers) and human participants. The foundational principles include:

Mixed-initiative control: Both the human and the machine contribute decisions and adapt their actions in response to each other’s input (Esfahani et al., 2017).
Feedback-driven adaptation: The optimization process incorporates direct human feedback—such as discrete choices, rankings, response times, or explicit constraints—at critical junctures, updating the model or search direction accordingly (Lage et al., 2018, Savage et al., 2023).
Preference learning: Human utility or preferences are modeled explicitly, often via statistical surrogates or latent functions in optimization frameworks, and refined through interaction (Schoinas et al., 31 Jan 2025, Ou et al., 2023).
Personalization: Human physiological, cognitive, or behavioral data are used to tailor system parameters on an individual basis, enabling adaptation to diverse users or operational contexts (Lagomarsino et al., 10 Sep 2024, Christou et al., 7 Oct 2025).

These principles are operationalized through looped interactions and critical interfaces that allow the human operator to shape the optimization trajectory.

2. Methodologies and Algorithmic Frameworks

HILO encompasses diverse methodological approaches depending on domain and system architecture:

Surrogate-based Bayesian Optimization: Probabilistic surrogate models (GPs or BNNs) fit to user preference or performance data, guide the selection of new candidate solutions. Acquisition functions (e.g., Expected Improvement, qEHVI, upper confidence bound) balance exploration and exploitation (Chan et al., 2022, Liao et al., 7 Mar 2025, Savage et al., 2023).
Multi-objective Optimization: Simultaneous optimization of conflicting objectives, such as utility and diversity (measured via covariance determinants), allows experts to select among Pareto-optimal candidate sets (Savage et al., 2023, Lagomarsino et al., 10 Sep 2024).
Evolutionary Strategies: Population-based algorithms such as CMA-ES adjust controller parameters (e.g., joint stiffness in exoskeletons) over repeated human trials, updating sample distributions in response to objective outcomes (Christou et al., 7 Oct 2025).
Closed-loop Control with Human Sensing: Real-time physiological and behavioral signals (e.g., attention, cognitive load, heart rate variability) are processed and fed into trajectory or safety zone adaptation modules for industrial robots (Lagomarsino et al., 10 Sep 2024).
Continual Learning Frameworks: Bayesian neural network surrogates are incrementally updated using generative replay from past users, blending population-level and individual-specific modeling to accelerate calibration (Liao et al., 7 Mar 2025).

Mathematical formulations and representative algorithms from leading HILO applications are displayed below:

Paper/Framework	Method	Key Formulation / Algorithmic Element
Bayesian Opt. (Savage et al., 2023)	Multi-objective BO + Expert Selection	maxₓ ( Σᵢ U(xᵢ),
Interpretability (Lage et al., 2018)	MAP with Human Prior	maxₘ p(X
Exoskeleton (Christou et al., 7 Oct 2025)	CMA-ES Human-in-the-Loop	Evolution path, sampling, objective minₖ[f₁J₁ + f₂J₂ + f₃J₃]
Robot Trajectory (Lagomarsino et al., 10 Sep 2024)	Real-time Multi-objective Control	minₕ (exec time, jerk), adaptive B-spline path, HRV feedback
Continual Optimization (Liao et al., 7 Mar 2025)	ConBO Bayesian Neural Network with Replay	aₜ(x) = wₘ,ₜ·EI_population(x) + (1−wₘ,ₜ)·EI_user(x)

Underpinning these frameworks is the formal modeling of both the objective functions (from physical metrics or human feedback) and the surrogate models that represent user preferences or capabilities.

3. Human Expertise, Judgment, and Feedback Integration

The integration of human capabilities in HILO occurs at multiple levels:

Selection and correction: Human operators select among discrete sets of system-generated alternatives, as optimizing continuous choices can be less reliable (Savage et al., 2023).
Preference modeling: Pairwise or tournament-based human choices are used to fit latent utility functions, often leveraging Gaussian processes or mixed-effects models (Schoinas et al., 31 Jan 2025).
Cognitive and physiological metrics: Human stress, attention, and workload are quantified and dynamically interpreted to adapt system safety, performance, and comfort. Camera-based head pose estimation and ECG-measured HRV provide real-time data (Lagomarsino et al., 10 Sep 2024).
Expertise moderation: The impact of user expertise is context-dependent: experts provide consistent, critical feedback that enhances objective outcome quality but report lower subjective satisfaction; novices converge rapidly to satisficing solutions (Ou et al., 2023).
Implicit versus explicit interaction: Systems may incorporate implicit feedback (observational or behavioral data) or explicit user interaction (direct ranking, selection, or constraint curation) (Christou et al., 7 Oct 2025, Jansen, 13 May 2025).

These mechanisms require sophisticated interface design to present alternatives, visualize feedback, and enable effective selection. Failure modes include preference inconsistency, cognitive biases (anchoring, loss aversion), and non-stationary human utility functions (Ou et al., 2022).

4. Practical Applications and Validation

HILO is widely applied across domains where personalization, adaptivity, or co-adaptation between system and user is critical:

Robotics and Manipulation: Haptic-guided master-slave robot grasping leverages operator’s grasp stability judgment, optimized via force cues from autonomous agents for task-relevant manipulability (Esfahani et al., 2017). Industrial collaborative robots tune trajectories in real time based on operator stress and attention (Lagomarsino et al., 10 Sep 2024).
Assistive and Medical Devices: Exoskeletons for gait training adapt joint parameters in response to human-in-the-loop feedback, though human behavioral variability may overwhelm measurable benefits (Christou et al., 7 Oct 2025).
Visual Prostheses: Retinal implant simulators incorporate user preference data to optimize stimulus encoding, with direct comparisons to simulated agents revealing human idiosyncrasies and optimization robustness (Schoinas et al., 31 Jan 2025).
User Interface Design: Inclusive design frameworks shift designers’ roles to constraint curation and algorithmic exploration, supporting scalable, individualized UI adaptation for accessibility (Jansen, 13 May 2025). Adaptive XR toolkits enable Pareto-optimization of reachability, visibility, and consistency for immersive applications.
Personalization and Continual Learning: Virtual reality keyboard calibration using CHiLO achieves faster adaptation times for new users by transferring population-level knowledge via generative replay in Bayesian neural network surrogates (Liao et al., 7 Mar 2025).
Motion Control: RL-based whole-body locomotion tracking in humanoid robots combines human-derived references with RL residual policies, stabilizing learning and achieving rapid adaptation across diverse tasks (Zhang et al., 5 Feb 2025).

Experimental validation ranges from usability studies (SUS scores for interactive systems (Liu et al., 2020)) to clinical phase evaluations and large-scale comparative analyses.

5. Challenges, Limitations, and Failure Modes

Several challenges confront the practical realization and evaluation of HILO systems:

Human variability and co-adaptation: Intra-individual fluctuations, learning effects, and adaptation to robotic assistance complicate performance attribution in personalized optimization (Christou et al., 7 Oct 2025).
Convergence failures: Optimization algorithms based on preferential choices may stall or oscillate due to inconsistent or contradictory human judgments, violating standard assumptions of utility stationarity and i.i.d. feedback (Ou et al., 2022).
Cognitive biases: Human evaluators are subject to anchoring, loss aversion, and context frame effects, which can bias optimization and degrade overall system convergence (Ou et al., 2022).
Expert agency and satisfaction: Automated guidance may limit designer expressiveness and perceived control, highlighting a trade-off between efficiency and creative ownership (Chan et al., 2022).
Scalability and sample efficiency: The expense of human trials motivates strategies to minimize the number of user studies (model-based acquisition selection (Lage et al., 2018), continual learning with generative replay (Liao et al., 7 Mar 2025)).
Constraint curation and transparency: As designer roles shift toward specifying search space constraints, questions arise regarding transparency, user agency, and ethical concerns, especially in inclusive and accessible design (Jansen, 13 May 2025).

Addressing these limitations requires hybrid methodologies, robust preference modeling, adaptive interfaces, and noise-resilient or therapist-integrated protocols.

6. Impact and Prospects

Human-in-the-Loop Optimisation advances solution relevance, safety, and user alignment in interactive, adaptive systems:

Performance gains: Empirical studies show improved outcome quality (e.g., HILO-stimuli outperforming naive or unrefined encoders (Schoinas et al., 31 Jan 2025), reduced operator stress and increased productivity in PRO-MIND (Lagomarsino et al., 10 Sep 2024)).
Personalization: Calibration protocols and continual learning frameworks show enhanced adaptation times and efficiency, especially with accumulating user bases (Liao et al., 7 Mar 2025).
Model interpretability: Direct human involvement in model selection refines interpretability proxies and enhances model trust and comprehension (Lage et al., 2018).
Broad application footprint: Domains ranging from robotics, medicine, user interface design, motion control, and optimization engineering benefit from HILO frameworks.
Future directions: Research focuses include integrating dynamic and feasibility constraints, mixed-initiative approaches for creative agency, adaptive methods sensitive to expertise and behavioral indicators, and scalable continual learning solutions for interactive systems.

A plausible implication is that as computational techniques mature and sensor fusion becomes standard, HILO systems will offer generalized, efficient, and personalized adaptation for diverse applications where automated modeling remains insufficient. The integration of robust preference modeling, adaptive interface design, and continual learning protocols will be essential to maximize the utility and acceptance of HILO systems.