Optimization algorithms are exceptionally good at exploring vast, high-dimensional search spaces, but they require a clear metric to evaluate success. When the target is subjective—such as musical rhythm, visual aesthetics, or game level design—the fitness function cannot be easily mathematical; it must be a human. This setup is known as Interactive Evolutionary Computation (IEC) or human-in-the-loop optimization.
However, a fundamental mismatch exists between machine optimization and human cognition. Algorithms typically require thousands of evaluations to converge, whereas humans experience rapid physical and mental exhaustion when manually scoring candidates over multiple generations. This bottleneck, known as human evaluation fatigue, causes user ratings to become noisy and inconsistent, ultimately stalling the creative process (Wang et al. in Improving the Performance of Predicting Users' Subjective Evaluation Characteristics to Reduce Their Fatigue in IEC). As noted by Chen et al. in Multi-objective Human-computer Co-operative Co-evolutionary Method Based on Non-dominated Sorting Strategy, this cognitive limit severely restricts the scale of problems that human-in-the-loop systems can solve, often discouraging users and slowing down creative workflows (Szerlip et al. in Demo: A Computer-Assisted Approach to Composing with MaestroGenesis).
To overcome this interactive bottleneck, researchers have developed several core strategies designed to minimize human cognitive load while maximizing optimization efficiency.
Constraining the Search Space via Generative Latent Spaces
One major driver of user fatigue is the "garbage filtering" problem. In traditional genetic algorithms, early generations consist of highly random, low-quality, or nonsensical candidates. Forcing a human to manually reject hundreds of broken designs before seeing anything interesting quickly drains their cognitive budget (Bontrager et al. in Deep Interactive Evolution).
To solve this, researchers shift the evolutionary search from raw parameter spaces to the latent spaces of pre-trained generative models. Instead of mutating individual pixels or raw structures, the algorithm evolves latent vectors that pass through a generator ensuring structural validity.
For instance, Togelius et al. in Optimizing Visual Properties of Game Content Through Neuroevolution use Compositional Pattern Producing Networks (CPPNs) to generate game content that satisfies functional physics constraints before any human interaction occurs. Similarly, in Interactive evolution and exploration within latent level-design space of generative adversarial networks, Schrum et al. allow users to navigate and blend the latent spaces of Generative Adversarial Networks (GANs) trained on existing video games. Because the underlying model restricts the search space to a manifold of coherent designs, the human evaluator only interacts with high-quality candidates, drastically reducing the fatigue associated with noisy initialization.
Approximating the User with Surrogate Models
If a human can only comfortably perform a few dozen evaluations, how can an algorithm perform the thousands of iterations necessary for convergence? The answer lies in building a computational proxy of the user: a surrogate model.
In Acoustical and behavioral heuristics for fast interactive sound design, Souaille et al. demonstrate that using surrogate models to prune the search space before a human ever evaluates a candidate allows users to reach optimal designs twice as fast. To make these surrogates effective with minimal data, Gabor et al. in Benchmarking Surrogate-Assisted Genetic Recommender Systems simulate a user's "mental model" using benchmark functions, proving that a surrogate can accurately evolve to match user preferences using very few true evaluations.
Furthermore, the way human feedback is processed matters. Wang et al. accelerate convergence by mapping relative human ratings to an absolute scale and prioritizing recent feedback, which accounts for the fact that human preferences shift during an interactive session. By combining this subjective feedback with objective feature metrics, systems can find personalized solutions in high-dimensional spaces with minimal user input, as shown by Horowitz in Generating Rhythms with Genetic Algorithms.
Intelligent Querying and Counterfactual Learning
Another strategy minimizes fatigue by optimizing when and how the system requests human input. Continuous, high-frequency feedback loops impose an unsustainable cognitive load on the trainer (Fang et al. in Multi-Agent Generative Adversarial Interactive Self-Imitation Learning for AUV Formation Control and Obstacle Avoidance).
Instead of demanding a rating for every single action or candidate generation, systems can use active learning to query the user strategically. In Active Reward Learning, Daniel et al. employ Bayesian Optimization to decide exactly when to ask for feedback, allowing a system to learn complex tasks with minimal human intervention.
When live testing is too slow or expensive, researchers can bypass the real-time human bottleneck entirely by learning from historical interaction logs. In Counterfactual Language Model Adaptation for Suggesting Phrases, Arnold et al. utilize counterfactual learning techniques to optimize phrase suggestion models using logged data. This eliminates the need for constant real-time human-in-the-loop experiments, which are traditionally slow and difficult to scale (Dzikovska et al. in Evaluating language understanding accuracy with respect to objective outcomes in a dialogue system).
Decentralizing and Automating Feedback Channels
Finally, user fatigue can be mitigated by changing the nature of the feedback channel itself—either by distributing the cognitive load across a crowd or by capturing implicit physiological responses.
In Aesthetic Bot: Interactively Evolving Game Maps on Twitter, Togelius et al. demonstrate that individual user fatigue can be entirely circumvented by offloading the evaluation of candidate solutions to a large, distributed audience via social media polls. This transforms a repetitive manual selection task into low-effort community engagement, allowing the system to train a neural network on thousands of public votes over extended periods.
Alternatively, the need for conscious evaluation can be removed altogether. In Augmented Iterations, Lotte et al. close the loop between human intuition and generative algorithms by monitoring real-time brain signals. By detecting the P300 "recognition" signal via EEG, the system automatically translates subconscious aesthetic relevance into a computational fitness score, allowing deep exploration of design spaces without the cognitive fatigue of manual sorting.
Key Foundational Work
Generating Rhythms with Genetic Algorithms (1994) — contributed 3 concept elements to this direction.
Improving the Performance of Predicting Users' Subjective Evaluation Characteristics to Reduce Their Fatigue in IEC (2005) — contributed 1 concept elements to this direction.