Evolutionary Surrogate-Assisted Prescription

• ESP is a framework that integrates surrogate predictors with evolutionary algorithms to efficiently optimize decision strategies in expensive evaluation environments.
• It relies on a cyclic process where historical data trains a predictor and an evolved prescriptor selects actions to maximize predicted outcomes.
• Applications span reinforcement learning, healthcare, and control systems, demonstrating improved sample efficiency and reduced real-world evaluation costs.

Evolutionary Surrogate-Assisted Prescription (ESP) is a framework for discovering decision strategies by combining a learned predictor or surrogate model with evolutionary search over a prescriptor that maps contexts to actions. In its general form, the predictor estimates the outcome of taking action $A$ in context $C$, and the prescriptor is evolved to choose actions that optimize those predicted outcomes (Francon et al., 2020). ESP belongs to the broader family of surrogate-assisted evolutionary algorithms for computationally expensive problems, but it centers prescription: the object being optimized is typically a policy, intervention strategy, or decision rule rather than only a static design vector. This suggests a shift from surrogate-assisted optimization of candidate solutions to surrogate-assisted optimization of actionable decision-making.

1. Definition and lineage

The immediate technical background of ESP is surrogate-assisted evolutionary optimization for expensive black-box objectives. That literature treats surrogate models as approximations to computationally expensive fitness functions and emphasizes five recurring questions: what type of fitness approximation to use, which approximation model to use, how to integrate the approximation model in the EA, how much approximation to use, and how to ensure reliable approximation (Bhattacharya, 2013). The surrogate repertoire discussed there includes polynomial models, Kriging, artificial neural networks, support vector machines, regression splines, radial basis function networks, and response surface models, together with model-management regimes such as fixed evolution control, adaptive evolution control, migration mechanisms, and active sampling (Bhattacharya, 2013).

Before ESP was formalized as a general prescription framework, surrogate-assisted evolutionary methods were already being used to optimize interventions and recommendations in application-specific settings. In a PhysiCell study on targeted delivery of a therapeutic compound to cancerous tumor cells, Gaussian process and multi-layer perceptron surrogate models were combined with a steady-state genetic algorithm to minimize the number of cancerous cells after simulated treatment; both model-assisted algorithms outperformed a standard evolutionary algorithm under a very small evaluation budget (Preen et al., 2018). In recommender systems, a surrogate-assisted interactive genetic algorithm was adapted so that a pool of user-evaluated items was used to construct an approximative model serving as a surrogate fitness function, and the model was updated after each batch of new evaluations so that it could evolve toward the user’s preferences (Gabor et al., 2019).

Against that background, ESP as named in the literature makes the prescription problem explicit. The 2020 framework defines a predictor $P_d$ and a prescriptor $P_s$, and extends the approach to sequential decision-making tasks in reinforcement learning benchmarks (Francon et al., 2020). Later reports use the same terminology for multi-objective control systems, including drinking water chlorination (Monsia et al., 26 Aug 2025). A plausible implication is that ESP is best understood not as a single algorithm, but as a design pattern: learn a surrogate on observed or simulated outcomes, evolve a prescriptor against that surrogate, and close the loop with periodic real or high-fidelity evaluation.

2. Formal framework and optimization loop

In the canonical ESP formulation, contexts or states belong to $\mathbb{C}$, actions belong to $\mathbb{A}$, and outcomes are denoted $O(C,A)$. The predictor is a supervised model trained to estimate outcomes,

$P_d(C,A)=O',$

and the prescriptor is a policy mapping contexts to actions,

$P_s(C)=A.$

The optimization objective is to evolve the prescriptor so that the predictor’s outputs are maximized over contexts; equivalently, the prescriptor searches over actions $A_i=P_s(C_i)$ to maximize predicted outcomes $C$0 across contexts and outcome dimensions (Francon et al., 2020).

The generic ESP loop has five steps. First, historical $C$1 data are used to train the predictor. Second, a population of prescriptors is evolved using the predictor as a fast evaluator. Third, the best prescriptor or prescriptors are tested in the real domain. Fourth, the new real-world data are added to the training set. Fifth, the predictor is retrained and the cycle repeats until convergence (Francon et al., 2020). The paper explicitly allows the predictor to be a random forest or a neural network trained with gradient descent, and the prescriptor is typically a neural network evolved by population-based evolutionary search (Francon et al., 2020).

For sequential decision-making, ESP reinterprets the same loop in reinforcement-learning terms. Contexts and actions occur in sequence, outcomes are future-discounted cumulative rewards, and the predictor estimates $C$2-values for $C$3 pairs. Elite prescriptors are run in the actual environment to obtain episodes, the predictor is retrained on the resulting trajectories, and neuroevolution continues against the updated surrogate (Francon et al., 2020). In this setting, the predictor is not merely a cheap proxy for an expensive function; it becomes a learned environment model through which policy search is regularized and accelerated.

3. Surrogate and search architectures

ESP does not prescribe a single surrogate class or a single evolutionary optimizer. The surrounding literature shows a wide range of surrogate architectures, acquisition rules, and evolutionary operators that are compatible with the framework.

In the cancer-treatment simulator study, the Gaussian process surrogate used an RBF kernel and produced both a mean prediction and a standard deviation for each candidate. Candidate selection then used the Expected Improvement acquisition function,

$C$4

with

$C$5

so that exploration and exploitation were balanced by uncertainty-aware pre-selection (Preen et al., 2018). The alternative surrogate in the same study was an MLP with 10 hidden neurons with rectified linear units, trained using the limited-memory Broyden–Fletcher–Goldfarb–Shanno algorithm; because the MLP did not output uncertainty, pre-selection was based solely on predicted mean fitness (Preen et al., 2018). The evolutionary core was a steady-state genetic algorithm with population size $C$6, tournament size $C$7, uniform crossover with probability $C$8, mutation rate $C$9 with $P_d$0, and step size $P_d$1 (Preen et al., 2018).

In the chlorination-control report, the prescriptor is evolved by NEAT, while NSGA-II handles reproduction and selection of NEAT genomes for multi-objective prescription. The surrogate is an LSTM-based neural network with two LSTMs—one for actions and one for observations—joined and followed by fully connected layers; it takes data from the past 10 time steps and is trained for 50 epochs per ESP iteration with Adam and early stopping (Monsia et al., 26 Aug 2025). The decomposition is explicit: the prescriptor $P_d$2 maps observations $P_d$3 to actions $P_d$4, while the predictor $P_d$5 maps observation-action pairs to rewards $P_d$6 (Monsia et al., 26 Aug 2025).

A more speculative extension appears in work on LLMs as surrogate models in evolutionary algorithms. There, surrogate-assisted selection is reformulated as prompt-based regression or classification, with the LLM performing zero-shot inference directly from historical evaluated solutions and candidate solutions, without model retraining. The resulting LLM-assisted EA selects promising offspring by lowest predicted value or “good” class and can optionally maintain a pool of unevaluated but promising candidates (Hao et al., 2024). This suggests that ESP-compatible surrogates need not always be retrained numerical regressors; they can also be inference-only models, provided they can rank or score candidate prescriptions.

4. Representative application domains

Representative ESP applications span reinforcement learning, pandemic intervention design, diabetes management, and water chlorination control (Francon et al., 2020, Miikkulainen et al., 2020, Khanna et al., 2024, Monsia et al., 26 Aug 2025).

Domain	Predictor and prescriptor	Reported behavior
Reinforcement learning	Predictor: random forest or neural network; prescriptor: evolved neural network	More sample efficient, lower variance, and lower regret than standard RL approaches
COVID-19 NPIs	Predictor: LSTM-based neural network; prescriptor: neural network evolved with NSGA-II	Pareto trade-offs between expected cases and total NPI stringency; school and workplace restrictions emerged as especially important
Diabetes management	Predictor: random forest; prescriptor: feedforward neural network	Pareto front reduced glucose deviation and bolus injections compared to the original data
Chlorination control	Predictor: LSTM-based neural network; prescriptor: NEAT with NSGA-II	Linear Pareto front between chlorine usage and violation bounds; only modest improvement on the two focused objectives

In the original ESP reinforcement-learning study, the majority of evaluations are done on the surrogate, and the authors report that ESP is more sample efficient, has lower variance, and lower regret than standard RL approaches. In a function-approximation domain, ESP converged to near-optimal performance in approximately 125 episodes, whereas DE and PPO failed to converge after 1000+ episodes. On Cart-Pole, ESP achieved perfect or near-perfect performance faster and more reliably than DE, PPO, and DQN; on Flappy Bird, with a random-forest predictor, ESP discovered a policy that solved the task at around 80,000 episodes, whereas DE failed even after an order of magnitude more training (Francon et al., 2020).

In pandemic control, ESP was used to prescribe non-pharmaceutical interventions over a 180-day horizon. The prescriptor consumed 21 days of case growth rates and output NPI levels for 8 interventions, while the predictor simulated expected case growth and a cost proxy defined by total stringency. The search used 250 networks per generation and NSGA-II to produce a Pareto front ranging from highly stringent to highly permissive strategies (Miikkulainen et al., 2020). The reported qualitative findings were that workplace and school restrictions were the most important and needed to be designed carefully, that lifting restrictions could be unreliable, and that evolution often discovered alternating or partial interventions that softened their societal cost (Miikkulainen et al., 2020).

In diabetes management, a random forest was trained on 729 hourly samples from 30 days of treatment and measurements of a single patient to predict deviation of blood glucose from the target value in the next hour, using 20 input features. A feedforward neural network with two fully connected layers, 16 hidden tanh units, and sigmoid outputs was then evolved to prescribe carbohydrates, basal pumping levels, and bolus injections. The search was multi-objective, with population 100 and 50 generations, and it produced a Pareto front that reduced deviation from the target and number of injections compared to the original data (Khanna et al., 2024).

In drinking water chlorination, ESP was applied to a real-world multi-objective control task from the IJCAI-2025 Drinking Water Chlorination Challenge. The implementation focused on two objectives—chlorine bound violations and cost of control—even though the task had five objectives in practice. ESP was run for four iterations; the best found solutions achieved cost control values in $P_d$7 and chlorine violation bounds in $P_d$8, and the 60 best solutions formed a linear Pareto front (Monsia et al., 26 Aug 2025). The same report also noted only modest improvement on the two focused objectives and little to no effect on the other three not directly targeted (Monsia et al., 26 Aug 2025).

5. Model management, portfolios, and adjacent surrogate-assisted advances

A central technical issue for ESP is model management: how to decide which candidates should be truly evaluated, which should be surrogate-evaluated, and how the surrogate should guide variation and selection. Adjacent work in surrogate-assisted evolutionary algorithms provides several mechanisms that are directly relevant.

One line of work treats algorithm choice itself as a prescription problem. For individual-based surrogate-assisted evolutionary algorithms, two portfolio frameworks were proposed for very expensive problems with a maximal number of fitness evaluations equal to $P_d$9: Par-IBSAEA, which runs all candidate algorithms in parallel with a shared database, and UCB-IBSAEA, which treats algorithms as arms in a multi-armed bandit and selects one per iteration using the UCB-Tuned policy (Tong et al., 2019). Candidate algorithms included EGO-LCB, VESAEA, and GORS-SSLPSO. The reported result was that both portfolio frameworks significantly outperformed any single algorithm on the benchmark suite and reduced selection risk; UCB-IBSAEA was especially strong on multimodal problems, while Par-IBSAEA was particularly strong on simpler unimodal functions (Tong et al., 2019). This suggests that, in ESP deployments where the problem characteristics are unknown and the evaluation budget is tight, automatic prescription may need to operate at two levels: prescribing actions within the domain and prescribing which surrogate-assisted optimizer should be trusted.

Another line of work argues that un-evaluated solutions may be valuable in expensive optimization. The proposed strategy retains high-quality, un-evaluated solutions predicted by the surrogate and lets them participate in the next generation’s reproduction as parents, so that the parent pool becomes $P_s$0 rather than only evaluated solutions $P_s$1. The method was instantiated for GA, DE, and EDA reproduction operators and was reported to outperform mainstream SAEAs and Bayesian optimization algorithms, while substantially improving computational efficiency in higher dimensions (Hao et al., 2024). For ESP, the relevant point is not only the reduction in expensive queries, but the claim that surrogate-selected but unconfirmed candidates can improve disparity in successive populations and thereby improve convergence.

Scaling to medium-scale expensive multi-objective problems has also been addressed through surrogate-specific transformations. A Gaussian-process-assisted EA for up to 50 decision variables used only variables correlated with each objective for surrogate construction, transformed an $P_s$2-objective problem into a $P_s$3-objective problem based on predicted mean and uncertainty, and selected batches of candidates for true evaluation by hypervolume contribution (Ruan et al., 2020). The method outperformed three state-of-the-art SAEAs in 95 out of 108 comparisons on ZDT and DTLZ benchmarks (Ruan et al., 2020). A plausible implication is that future ESP systems with large state-action representations may benefit from objective-wise variable selection and explicit exploration-exploitation decomposition rather than a single monolithic surrogate.

For genetic programming in dynamic scheduling, surrogate-assisted GP has required problem-specific phenotypic characterization. In the DMRCPSP study, a rank-based PC vector was built from decision situations involving ordering of eligible activity-mode pairs and activity groups; surrogate fitness prediction then used Manhattan distance and nearest-neighbor transfer from already evaluated GP individuals. The resulting surrogate-assisted GP identified high-quality heuristic rules consistently earlier than the state-of-the-art GP approach while introducing only marginal computational overhead (Tian et al., 17 Mar 2026). This suggests that when ESP prescriptors are symbolic or rule-based rather than neural, behavior-level characterization may be more informative than genotype-level similarity.

6. Limitations, misconceptions, and research directions

A recurrent misconception is that surrogate assistance automatically improves optimization. The expensive-optimization literature argues the opposite: surrogates can introduce false optima, cause incorrect convergence, and fail if model management is poor. The recommended practices are to combine surrogate-based evaluations with periodic true evaluations, use adaptive evolution control, update the surrogate online as new true evaluations become available, and employ active learning or uncertainty quantification to decide where to query the true function (Bhattacharya, 2013). In that sense, ESP is not a substitute for real evaluation; it is a framework for rationing real evaluation.

Empirical studies also report mixed outcomes when surrogate trust is not well calibrated. In the batch processing problem, surrogate assistance was implemented through the Probablistic Surrogate-Assisted Framework, wrapping GA and DE into PSAF-GA and PSAF-DE. The reported outcome was conditional: surrogate assistance often improved AESR and AGSR, especially for longer horizons, but in other time horizons it maintained the solutions or showed some deterioration, and the study highlighted the need to tune the hyper-parameters used by the surrogate-assisted framework (Variawa et al., 2022). This is directly relevant to ESP because prescriptive quality depends not only on predictor accuracy, but also on the choice of tournament size, surrogate generations, infill behavior, and audit frequency.

Another limitation is partial objective coverage. In chlorination control, only 2 of the 5 competition objectives were included in the evolved reward, and the lower bound violations dominated the learning signal. The authors explicitly note that this limited policy generality and real-world applicability, and they proposed curriculum learning and more diverse initializations as remedies (Monsia et al., 26 Aug 2025). The broader lesson is that surrogate-mediated prescription can optimize the objectives it is given while neglecting omitted objectives or objectives overwhelmed by dominant penalties.

A further misconception is that any expressive model is automatically a good surrogate. Work on LLM-assisted evolutionary algorithms reports that zero-shot LLM surrogates can achieve performance comparable to traditional surrogate models using only inference, particularly on low to moderate dimensional tasks, but it also notes that inference time is acceptable but not trivial and that performance is expected to degrade on much higher dimensions (Hao et al., 2024). For purely numerical tasks, lightweight regressors may still be preferable. Likewise, the original ESP paper attributes some of ESP’s empirical gains to regularization through imperfect surrogates, but that regularization is beneficial only insofar as the predictor smooths a deceptive landscape without destroying the signal needed for good prescription (Francon et al., 2020).

Taken together, the literature portrays ESP as a technically flexible but operationally delicate framework. Its central promise is efficient search for decision strategies in settings where direct evaluation is expensive, risky, or slow; its central challenge is ensuring that the surrogate, the evolutionary dynamics, and the objective formulation remain aligned closely enough that the prescribed strategy transfers from the learned model to the real system.