A Policy–Machine Learning Hybrid Approach to Evaluate Trap Mesh Selectivity: A Case Study on Pseudopleuronectes yokohamae

Abstract

A machine learning-based policy–utility framework was developed to assess trap mesh sizes (35–80 mm) in the Marbled Flounder fishery and reframe traditional selectivity analysis into a policy-oriented decision context. A utility function integrating catch per unit effort (CPUE), the immature proportion, and the bycatch ratio was constructed from experimental data collected in 2015–2016 and assessed under multiple policy weighting scenarios. Gradient boosting models trained on the 2016 data and validated with the 2015 data demonstrated strong predictive accuracy. The empirically optimized weighting set (α* = 0.79, β* = 2.36, and γ* = 0.79) produced high agreement between predicted and observed utilities (root mean square error ≈ 0.22; r = 0.901). Variable importance analysis identified the immature proportion as the main driver of utility variation; bycatch ratio and CPUE made smaller contributions. Scenario-based simulations showed a shift in the optimal mesh size, from 65 mm in 2015 to 80 mm in 2016, that reflects interannual changes to population size structure and bycatch composition. Policy regret analysis (comparing 65 mm to 80 mm) indicated consistently low regret (ΔU ≈ 0.12–0.15) and relative regret (<80%) values. This integrated utility–regret framework provides a dynamic, policy-relevant tool for linking trap selectivity information to management objectives.

1. Introduction

The ecological impact of pot fishing gear is generally considered to be less severe than that of bottom trawls and bottom-set gillnets [1,2]. Pots are well recognized for their advantages (e.g., species selectivity and size selectivity), and catches are typically retained in good condition [3]. Despite their relatively low habitat impact, pots can still cause benthic disturbances when they contact the seabed or are dragged along it; the scale of the disturbance depends on the pot size, weight, material, placement method, and retrieval speed [1]. Data on pot fisheries bycatch rates is limited or missing due to the lack of adequate observer coverage. Evidence from regions with monitoring suggests that pot bycatch is often dominated by immature individuals of the target species [1,4]. Although most pot fisheries in Canada have little or no observer coverage, the sablefish fishery in British Columbia has been subject to 100% electronic monitoring since 2006. Some snow crab fisheries in Nova Scotia have reached observer coverage levels of up to 30%. In the sablefish fishery, approximately 8% of immature individuals were discarded, while the snow crab fishery reported the very low discard rate of 0.01% [1].

Previous studies evaluated the suitability of mesh size regulations as a means of conserving and rebuilding fishery resources in pot fisheries [4,5,6,7,8,9,10,11]. These studies assessed gear selectivity (i.e., capturing or excluding individuals of specific species or sizes) by modifying mesh size [5,6,7,8,11] or incorporating escape devices [4,9,10]. Selectivity analysis was typically conducted by fitting a logistic SELECT model [12] to estimate selectivity curves and derive important parameters, such as 50% retention length (L50) and selection range (SR). However, these approaches are largely restricted to assessments of the physical selectivity of individual gears. They offer limited capacity to incorporate broader considerations, such as operational efficiency, the impact of bycatch, and policy relevance. SELECT-based analyses are valuable for quantifying biological selectivity, but reconfiguring them as policy-relevant, integrated management indices requires an analytical framework that simultaneously accounts for ecological, economic, and operational factors.

Recently, Koo and Kwon [11,13] expanded a traditional selectivity analysis of integrated ecological economic assessments by applying machine learning techniques to determine optimal mesh sizes for pot and gillnet fisheries based on biological factors (e.g., length at maturity and bycatch ratio) and economic indicators (e.g., catch per unit effort [CPUE]). However, these studies focused on evaluating the efficiency of individual gears and therefore have a limited ability to assess broader policy effectiveness or management performance.

Several policy-oriented frameworks have been developed to support decision-making in fisheries management. Zhou et al. [14] proposed a data-limited ecological risk assessment (ERA) framework that can quantify cumulative fishing risks to bycatch species and help establish management priorities, even in data-poor fisheries. Bastardie et al. [15] examined the resilience of the European Union Common Fisheries Policy (CFP) to climate change and fuel costs by developing a decision-making framework that integrates ecological, economic, and operational trade-offs. Kwon et al. [2] developed the Gear-Based Fisheries Management Index (GFMI), a policy implementation tool that evaluates gear controllability, environmental sustainability, and operational functionality. Although these studies did not directly analyze gear selectivity, they offer scientific foundations for structuring policy decisions and setting management priorities in practical fishing contexts.

Building on these policy-oriented frameworks, the present study reframes mesh size regulation, a classic selectivity management issue, from a policy–utility perspective. We focus on the Marbled Flounder (Pseudopleuronectes yokohamae) pot fishery, which has high economic value in Republic of Korea [16,17]. Four policy scenarios were developed: a productivity-oriented scenario emphasizing fishing efficiency and yield; a conservation-oriented scenario prioritizing resource protection and bycatch reduction; a balanced scenario assigning equal importance to productivity and conservation goals; and an empirically optimal scenario defined by the weight combination (α, β, and γ) derived through cross-validation. For each scenario, a policy–utility index (Policy_U) was calculated using different weights (α, β, and γ) for CPUE, the immature proportion, and the bycatch rate. A gradient boosting machine learning model was then used to analyze the sensitivity of utility outcomes and derive optimal mesh sizes under each scenario. Building on traditional selectivity analysis, this approach transcends traditional selectivity analysis to provide a dynamic analytical framework capable of evaluating how fishing efficiency and ecological sustainability change under different policy objectives.

2. Materials and Methods

2.1. Field Experiment and Data Collection

This experiment was conducted in the coastal waters of southern Republic of Korea from 2015 to 2016 (Figure 1). Cylindrical traps were used to target Marbled Flounder. Each trap was constructed of 21-ply polyethylene (PE) mesh with an upper diameter of 940 mm, a lower diameter of 820 mm, and a height of 220 mm. The frame, made of 11–13 mm rebar, had two oval entrances (one on each side; 150 mm high and 220 mm wide) reinforced with 1.6 mm stainless steel wire (Figure 2).

Traps with mesh sizes of 35, 50, 65, and 80 mm were used in both years (Figure 3), and the 2016 experiment included an additional mesh size of 100 mm. However, the 100 mm mesh was excluded from this study because its catch proportion was extremely low (approximately 2% across 2015–2016). As a result, the selectivity curve had an unstable fit [11].

A total of eight experimental fishing trips were carried out in 2015 (all in May) and 24 trips in 2016 (three in March, ten in April, six in May, and three in June). Fishing effort was concentrated in April and May, when market demand for Marbled Flounder is highest. A total of 50 traps were deployed each trip in five sets of ten traps with five different mesh sizes (each mesh size was duplicated). Traps were baited with chopped anchovies (Engraulis japonicus) or crushed oysters and soaked for approximately 12 h before retrieval. To minimize spatial bias, the order of mesh sizes was alternated, and traps were positioned at intervals of approximately 15 m (Figure 3).

All captured specimens were identified to the species level and counted. For Marbled Flounder, total length (TL) and body height (BH) were measured to the nearest 0.1 cm using a measuring board, and body weight was measured to the nearest 1 g using an electronic scale (CAS SW-02; CAS, Yangju-si, Gyeonggi-do, Republic of Korea). Individuals with a total length ≥ 20.0 cm were classified as mature based on the legal maturity standards for trap fisheries in Republic of Korea.

2.2. Data Analysis and Modeling Framework

To quantitatively evaluate the trade-off between catch efficiency and resource conservation in the Marbled Flounder trap fishery, a policy–utility machine learning framework was developed. This approach complements the traditional logistic selectivity curve analysis and is designed to identify the optimal mesh size under different policy objectives, such as productivity, conservation, and bycatch reduction, using experimental fishing data.

2.2.1. Definition of the Policy–Utility Function

For each fishing unit (year–month–operation–mesh size), policy–utility (U) was calculated using the following equation:

$$\mathbf{P}\mathbf{o}\mathbf{l}\mathbf{i}\mathbf{c}\mathbf{y}-\mathbf{U}=\mathsf{\alpha }\left(\mathbf{C}\mathbf{P}\mathbf{U}\mathbf{E}\right)-\mathsf{\beta }\left(\mathbf{I}\mathbf{m}\mathbf{m}\mathbf{a}\mathbf{t}\mathbf{u}\mathbf{r}\mathbf{e} \mathbf{r}\mathbf{a}\mathbf{t}\mathbf{e}\right)-\mathsf{\gamma }\left(\mathbf{B}\mathbf{y}\mathbf{c}\mathbf{a}\mathbf{t}\mathbf{c}\mathbf{h} \mathbf{r}\mathbf{a}\mathbf{t}\mathbf{e}\right)$$

(1)

Here, “CPUE” is catch per unit effort (a productivity indicator), “Immature rate” denotes the proportion of individuals below the legal mature length of 20 cm (a conservation indicator), and “Bycatch rate” represents the proportion of non-target species (an indicator of bycatch reduction). The parameters “α”, “β”, and “γ” are weighting factors that assign the relative importance of each policy objective. The directional effects of these variables on the overall policy–utility function are summarized in

Table 1. Directional effects of each variable on the policy–utility function.

Variable	Direction	Description
CPUE	Positive ($+$)	Positively contributes to productivity and fishing efficiency.
Immature proportion	Negative ($-$)	Negatively associated with resource sustainability due to immature catch.
Bycatch rate	Negative ($-$)	Represents ecological and economic penalties of non-target catch.

.

2.2.2. Data Preprocessing

CPUE, bycatch rate, and immature proportion (<20 cm total length) were calculated for each fishing unit (defined by year, month, experimental plot, and mesh size) using the experimental data. To capture temporal variability during the experimental period, the four-month window from March to June was treated as a single annual cycle, and monthly data were converted into a sinusoidal format. This allowed the model to recognize nonlinear temporal patterns throughout the fishing season (e.g., increasing water temperature or seasonal shifts in bycatch composition).

Monthly variability was encoded using sine–cosine terms (m-sinusoidal and m-cosine) based on the March–June pattern from the 2016 training data. Since the 2015 external validation data was collected during a single month (May), the sin–cos terms were treated as constants for that dataset.

To prevent data from being split across training and validation folds for the same fishing area and month, a grouping identifier (group_key) was created using year–month–experimental plot combinations.

2.2.3. Training and Validation Data Composition

Although eight experimental hauls were conducted in 2015, the number of flatfish caught was small, which made the data unsuitable for training a standalone model. Therefore, the Policy-U model was trained solely on the 2016 data, and the 2015 observations were used only for external validation to assess generalization performance.

Internal validation of the 2016 dataset was conducted using a group k-fold cross-validation analysis (k = 3). To reduce spatial and temporal dependence, observations from the same month and experimental plot were not split across multiple folds.

2.2.4. Simulations of Optimal Mesh Sizes Under Different Policy Scenarios

Four policy scenarios were developed to reflect different management objectives (

Table 2. Policy weighting scenarios and corresponding characteristics.

Scenario Type	Weight Combination (α, β, γ)	Characteristic
Productivity-oriented	1.5, 0.5, 0.5	Emphasis on productivity
Balanced	1.0, 1.0, 1.0	Policy-neutral
Conservation-oriented	0.4, 1.3, 1.3	Emphasis on conservation & bycatch reduction
Optimal	(α*, β*, γ*)	Empirical equilibrium point

). The productivity-focused scenario places the greatest emphasis on CPUE, while the conservation-focused scenario assigns higher weights to reducing bycatch and the immature catch. The balanced scenario gives equal weight to all three factors. Finally, the empirically optimal scenario represents the set of weights (α*, β*, and γ*) derived from the cross-validation of the 2016 dataset.

Candidate values for α, β, and γ were set to 0.5, 1.0, and 1.5, respectively. All 27 possible combinations were evaluated. For each combination, the policy–utility value was calculated and applied to the training data. The model was trained using a gradient boosting regressor, and three-fold grouped cross-validation was performed to calculate the scale-invariant root mean square error (RMSE) between predicted and observed utilities.

The combination of weighting factors that yielded the lowest RMSE was selected to provide the empirically optimal weighting values (α*, β*, and γ*). Finally, the selected weights were normalized to sum to 3 to ensure comparability across scenarios.

2.2.5. Model Validation and Policy Regret Analysis

The model trained on the 2016 dataset was externally validated using the 2015 data. Its predictive performance was evaluated through Pearson’s correlation coefficient (r) and the RMSE between the predicted and the observed utility values. After calculating U for each observation using the optimal weight combination (i.e., α, β, and γ), the input variables (X) and output variable (Y) were defined accordingly.

To assess the long-term performance of various mesh size policies under conditions where adaptive ecological and operational responses may occur, this study adopted the concept of policy regret, which was introduced by Arora et al. [18]. Policy regret is a comparison of outcomes under two policies that evaluates policy performance and accounts for the reaction of the environment (e.g., catch composition, bycatch levels, and the immature proportion) to policy choices. This approach allows the identification of conditions that may result in regret for a given policy.

In this study, policy regret was evaluated when the mesh size that maximized utility in the training data differed from the mesh size that maximized utility in the validation data. In these scenarios, the magnitude of regret represents the degree to which the initially chosen policy may underperform because of interannual changes in population structure or bycatch rate.

For example, we examined whether differences in predicted utility (ΔU) arose when an alternative mesh size policy (e.g., 65 mm) was maintained instead of the baseline policy (e.g., 80 mm), to reflect reactive changes in CPUE, bycatch rate, and immature fish proportion.

Policy regret was defined as the difference in predicted mean utility between two competing mesh size policies, expressed as:

$$∆\mathit{U}={\mathit{U}}_{\mathit{a}\mathit{l}\mathit{t}\mathit{e}\mathit{r}\mathit{n}\mathit{a}\mathit{t}\mathit{i}\mathit{v}\mathit{e}}-{\mathit{U}}_{\mathit{b}\mathit{a}\mathit{s}\mathit{e}\mathit{l}\mathit{i}\mathit{n}\mathit{e}}$$

(2)

Here, “U” represents the policy–utility estimated using the gradient boosting model. The primary predictors consisted of CPUE, the immature fish proportion, and the bycatch rate.

Bootstrapping (10,000 resamples) was used to estimate the 95% confidence intervals (CI). Cohen’s d and the relative regret (ΔU/U range) were also calculated.

3. Results

3.1. Model Validation and Interpretation

The analysis of the 27 possible combinations of utility weights revealed that the lowest average RMSE (0.1354) was achieved when α* = 0.5, β* = 1.5, and γ* = 0.5. This result indicates that placing less emphasis on productivity (α) and assigning a stronger penalty to the immature fish ratio (β) provides the best balance between predictive stability and ecological selectivity.

After normalizing the empirically optimal weights so that their sum equaled 3, the final optimal values were α* = 0.79, β* = 2.36, and γ* = 0.79. These weights were used in the machine learning-based utility prediction model and subsequent simulated scenarios. Internal cross-validation results showed scale-invariant RMSE values of 0.0768, 0.1719, and 0.1554 across the three folds, with an average RMSE of 0.1347 ± 0.048.

External validation was conducted using an independent dataset from 2015. The Pearson correlation coefficient between predicted and observed utility values was r = 0.901 (p < 0.001), which demonstrates the model’s strong predictive consistency, and the RMSE was 0.2235. These results confirm that the model trained under the empirically optimal weights (α* = 0.79, β* = 2.36, and γ* = 0.79) retains a stable predictive performance despite interannual variability. Feature importance analysis (Figure 4) indicated that the immature fish proportion was the most influential predictor. It substantially exceeded the contributions of other variables. Bycatch rate had the second-largest influence, while CPUE (productivity) exhibited relatively low explanatory power. In contrast, the seasonal m-sinusoidal terms, m-cosine terms, and mesh size had minimal effects on model predictions.

3.2. Scenario-Specific Utility and Selection Pattern Analysis

Using data from 2015 and 2016, we calculated the average difference in utility (ΔU) by month and mesh size using four policy weight combinations of α, β, and γ. The results revealed marked differences in utility levels and fluctuations across years (Figure 5 and Figure 6).

In 2015, utility remained stable at a high overall level. The average ΔU in the productivity-focused scenario was approximately −0.5, and there were no significant differences across the mesh sections (Figure 5a). In the balanced scenario, ΔU decreased slightly to −0.9, but the fluctuations were small (Figure 5b). In the conservation-focused scenario, ΔU was the lowest, at −1.3, and there was a pronounced decrease in utility in the 35–50 mm range (Figure 5c). The empirically optimal scenario, which maintained a balance between productivity and conservation, had ΔUs ranging from −0.55 to −0.75 and the highest utilities at 65–80 mm (Figure 5d).

In 2016, utility values decreased across all scenarios and fluctuated widely (Figure 6). The productivity-oriented ΔU ranged from −0.2 to −0.6. Small and medium-sized meshes demonstrated lower utilities (Figure 6a). The equilibrium ΔU ranged from −0.6 to −1.2, and utility decreased more rapidly with decreasing mesh sizes (Figure 6b). The conservation-oriented ΔU ranged from −1.5 or less, and the lowest utility occurred in the 35–50 mm mesh size range in the May–June period (Figure 6c). The empirically optimal ΔU ranged from −0.7 to −1.6, which was slightly higher than the conservation-oriented ΔU, and utility was moderated in the 65–80 mm range (Figure 6d).

Figure 7 and Figure 8 present the mesh sizes that maximized policy utility on a monthly basis for 2015 and 2016, respectively, with 65 mm identified as the dominant optimal mesh size in 2015 and 80 mm in 2016. In these figures, the optimal mesh size under each policy scenario is indicated by the largest marker (green), whereas smaller markers (blue) represent other tested mesh sizes.

3.3. Policy Regret Analysis

To assess policy robustness, we calculated policy regret (i.e., ∆U = U₆₅ − U₈₀) as the difference in utility between the 80 mm mesh sizes (i.e., the optimal mesh size derived from the training data in 2016) and the validation data (2015). The average ∆U across all scenarios was small, ranging from 0.12 to 0.15, which indicates just a small difference in utility between the 65 mm and 80 mm mesh sizes (

Table 3. Comparison of mean difference in utility (ΔU) and relative regret among four policy weighting scenarios.

Scenario	∆U_mean	CI_low_95	CI_high_95	Relative_regret_pct
Productivity	0.119475	0.119475	0.119475	100
Balanced	0.132458	0.132458	0.132458	100
Conservation	0.142195	0.142195	0.142195	77.14551
Empirically optimal	0.148676	0.148676	0.148676	72.24447

).

The productivity-focused (∆U = 0.119) and balanced (∆U = 0.132) scenarios showed the lowest regret values, indicating that the large mesh size (80 mm) exhibited high policy consistency. In the conservation-focused (ΔU = 0.142) and empirically optimal (ΔU = 0.149) scenarios, the 65 mm mesh showed somewhat higher utility, and the relative regret rates were approximately 77.1% and 72.2%, respectively. Both the utility and regret rates fell within the criteria (i.e., ΔU < 0.15, regret rate < 80%) of a robust policy. In other words, policy regret was minimal for all scenarios, and the 80 mm mesh was a stable, optimal choice across years.

4. Discussion

This study used a machine learning-based utility model to estimate policy–utility (ΔU) by integrating productivity (α), the immature fish proportion (β), and the bycatch rate (γ). The validation results showed strong stability (r = 0.901; RMSE = 0.22 on independent 2015–2016 data), demonstrating that the proposed model can reliably capture biological selectivity and the economic and ecological trade-offs that occur in the fishing industry.

However, the analysis data used in this study were collected in 2015–2016, approximately ten years prior to the present analysis. Because fishery resource status, population size structure, and environmental conditions can change over time, the optimal mesh size derived from this study must be interpreted within the context of the time when the data were collected.

Accordingly, the primary contribution of this study is not the identification of a time-invariant “optimal” mesh size, but the presentation of an analytical framework that can be explicitly recalibrated as new field data become available. Recalibration entails updating the empirical distributions of CPUE, juvenile fish proportion, and bycatch rate using recent observations, followed by re-estimation of the policy utility function and retraining of the machine learning model under an unchanged weighting structure. Since the structure of the utility function and the policy regret evaluation method are maintained, mesh recommendations derived from updated data can be compared with previous results, and the impact of changing ecological or operational conditions can be assessed based on data rather than assumptions.

The variable importance analysis revealed that the proportion of immature fish was the dominant driver of utility variation, which suggests that the key determinant of ΔU was not the mesh size per se but the proportion of immature fish retained by the gear. CPUE and bycatch rate showed complementary effects on utility, but seasonal factors and mesh size exhibited limited direct influence.

The long-term risks associated with mesh size choices were evaluated through policy regret analysis. In the productivity-focused scenario, 65 mm emerged as the optimal mesh size (i.e., high utility and low regret) while the 65–80 mm range provided the most favorable trade-offs in the conservation and balanced-policy scenarios. Mesh sizes below 50 mm were identified as high-risk options due to elevated regret arising from the excessive capture of immature individuals. Although mesh sizes above 80 mm are ecologically advantageous, they tended to increase regret levels because of reduced short-term productivity.

Furthermore, the model in this study was trained and validated using data from two years, and it is important to note that the 2015 data, in particular, has a relatively small sample size. This limited time span may limit the ability to sufficiently capture natural interannual fluctuations in population dynamics, environmental variability, and fishing conditions. To address this, this study employed grouped cross-validation, year-based external validation, and policy regret analysis to assess robustness under data constraints. These procedures are not intended to comprehensively reproduce long-term ecological variability, but rather to assess whether the derived management results are stable despite the limited sample size.

Previous studies on trap selectivity have largely focused on quantifying the single-factor effects of mesh size or structural modifications on catch characteristics, such as length distribution, bycatch rate, and discard mortality (

Table 4. Comparison of previous studies on fishing gear selectivity and management frameworks in relation to the present study.

Study	Approach	Strength	Limitation	Relevance to Present Study
Koo and Kwon [11]	SELECT + Decision Tree	Combines biology + economics	Scenario-limited maturity thresholds	Baseline for ΔU modeling
Broadhurst et al. [19]	GLMM + GAM	Shows cumulative effects of mesh + gaps	Species/site-specific	Gear modification reference for scenarios
Jeong et al. [5]	Extended SELECT	Detailed size-selectivity curves	Control-gear bias	Provides biological selectivity scaling
Rudershausen et al. [20]	Proportional selectivity + mortality	Links selectivity to survival	Short-term focus, habitat-specific	Motivates multi-factor inputs
Present study	GBM(Gradient Boosting Machine)-based ΔU + policy regret	Multi-objective optimization	Complex model	Integrates biology, economy, and policy

). Approaches used to evaluate effects include the extended SELECT model [4], Poisson Generalized Linear Mixed Model (GLMM)/Generalized Additive Model (GAM) [19], and proportional logistic selectivity models [20]. Although these methods are effective for evaluating specific structural changes, they offer a partial representation of the complex ecological interactions that occur in practical fishing operations, where CPUE, the immature rate, bycatch rate, and seasonal conditions can change simultaneously.

Using the same experimental dataset as Koo and Kwon [11], but applying a different analytical method, the present study also identified 80 mm as the optimal mesh size when adopting a length at maturity threshold of 20 cm. Koo and Kwon [11] combined a SELECT-based retention model with a decision tree classifier to estimate mesh-specific retention curves. However, in their framework, CPUE, the immature fraction, and the bycatch fraction were treated solely as dependent outcomes of mesh size. Interactions among these factors were not considered, and performance across different policy objectives or years was not evaluated.

Building on conventional selectivity-based approaches, this study incorporated management objectives (i.e., productivity, decreasing the immature catch, and bycatch reduction) into a single ΔU utility function that models interactions among key variables using machine learning and evaluates robustness through cross-year external validation and policy regret analysis. The fact that both analytical approaches reached the same conclusion, an optimal mesh size of 80 mm, indicates that this result is not driven by methodological choice but by the ecological and economic structure of the dataset. This consistency reinforces the conclusion that an 80 mm mesh size represents a defensible management option that balances ecological conservation with economic efficiency.

Furthermore, the ΔU-based decision framework presented here extends the interpretative scope of conventional selectivity-based studies because it explicitly incorporates policy goals to compare alternative management scenarios and assess risk through regret analysis. As a result, it provides a level of managerial relevance and analytical depth that complement, earlier selectivity-focused studies were not designed to deliver.

Nonetheless, it should be emphasized that this framework was developed using data from a single species and a single marine area (Goseong Bay, Republic of Korea). Because population parameters, community structure, and environmental drivers can vary substantially among regions, the results of this study should not be interpreted as universally applicable to all marbled flounder pot fisheries. Rather than being proposed as a universal regulatory standard, this framework is presented as a site-calibrated analytical tool that supports region-specific decision-making through reparameterization using locally relevant data from the target fishery or region.

Policy regret analysis indicated that the utility difference between the two candidate mesh sizes (65 mm vs. 80 mm) was minimal. The calculated regret values (ΔU = U₆₅ − U₈₀) ranged from 0.12 to 0.15 across all policy weighting scenarios, and the corresponding relative regret (72–77%) remained well below the robustness threshold. These findings demonstrate that 80 mm represents a stable, robust, and optimal policy choice across years, with utility differences that fall within the range of statistical uncertainty.

Even in the empirically optimal scenario, where a mesh size of 65 mm yielded a marginally higher ΔU, the magnitude of the loss associated with selecting 80 mm instead of 65 mm was small enough to be acceptable in a policy context. This stability implies that a mesh size of 80 mm provides consistent performance under productivity-focused and conservation-oriented objectives, and these results support its suitability as an interannual management strategy. Overall, the low regret levels suggest that the larger mesh size maintains policy consistency and ensures a balance between ecological protection and economic efficiency, even under varying environmental or operational conditions.

A number of efforts in the fisheries sector have sought to incorporate uncertainty and regret-based decision-making into policy design. Stohs and Harmon [21] applied a Bayesian posterior predictive framework to evaluate experimental deep-set buoy gear and assessed policy performance as the probability that the bycatch of protected species would exceed management thresholds. They used a risk-based evaluation that relied on exceedance probabilities rather than simple average bycatch rates. Holzer [22] developed a game-theoretic model of fishers’ choices between a derby-style common-pool fishery and an Individual Transferable Quota (ITQ) regime. Using the Maryland striped bass ITQ transition, the study demonstrated that allowing individual harvesters to voluntarily enter a property rights system can reduce political resistance and lower the transaction costs associated with reform. Koemle et al. [23] employed a latent class choice model to jointly estimate random utility maximization (RUM) and random regret minimization (RRM) decision rules in a recreational pike fishery. They demonstrated that anglers with a higher involvement in fishing are more likely to follow RRM rules and favor conservation-oriented regulations. Their results also indicated that welfare estimates based on identical policy changes can differ substantially depending on whether RUM or RRM is used to interpret angler behavior.

In comparison, the present study adopts a micro-technical, design-focused perspective. Concentrating on the internal configuration of a single commercial trap and its mesh size, we used machine learning methods to estimate an integrated biological–economic utility index that incorporates CPUE, juvenile fish protection, and bycatch reduction. Policy regret (ΔU) was used to evaluate the robustness of alternative mesh sizes across years and policy weighting scenarios.

Together, these studies represent a continuum of decision support tools for fisheries management that describe risk-based assessments of new gear technologies [21], strategic analyses of fisher incentives and regime choice [22], behavioral models capturing heterogeneous decision rules among resource users [23], and fine-scale gear-design optimization with regret-based robustness diagnostics (this study). This body of work collectively suggests that effective mesh size regulation requires the explicit quantification of biological and economic uncertainty integrated with a consideration of fisher behavior and incentives and a regret-based evaluation of whether “nominally optimal” designs remain acceptable as management objectives, years, and decision-making rules evolve. The findings derived from our policy scenario evaluation and policy regret analysis suggest that this study provides a practical decision support tool that complements existing approaches by enabling assessments of interannual stability in mesh size regulations and the robustness of management outcomes under shifting policy objectives (i.e., productivity versus conservation) within an ecosystem-based fisheries management framework.

This study has several limitations. First, the analysis relied on flounder trap data collected in Goseong Bay during 2015–2016, which restricts broader geographic and species-level generalizations. The short time frame also limits the model’s ability to capture long-term ecological fluctuations in the proportion of juveniles, bycatch composition, and fishing performance. Future research should incorporate multi-regional, long-term datasets and population dynamics models to improve ecological relevance.

Second, although the gradient boosting model demonstrated strong predictive accuracy, its nonlinear structure limits interpretability. Applying explainable AI techniques, such as SHapley Additive exPlanations (SHAP) values or partial dependence plots, would enhance its clarity and coherence. Likewise, it would be useful to develop quantitative linkages between ΔU or policy regret metrics and the reference points described by the Sustainability Assessment for Fishing Effects (SAFE) or the Ecological Risk Assessment for the Effects of Fishing (ERAEF).

Third, the included economic and social factors were simplified. CPUE was used as a proxy for productivity, but it did not account for management costs, labor requirements, market variability, or social acceptance. An expanded utility function that includes these components would provide more realistic policy assessments.

Accordingly, the “optimal” mesh size derived in this study should not be interpreted as a full bioeconomic optimum, but rather as an ecological–operational optimum under simplified economic assumptions. Incorporating price volatility, weight—class—dependent market values, and operational costs into the utility function represents an important direction for future research.

Finally, since South Korea’s coastal and offshore fisheries operate in a multi-species, multi-gear system, extending the model to consider interactions among gear types (e.g., competition, effort redistribution, and opportunity costs) would substantially increase the practical applicability of the integrated policy framework.

5. Conclusions

This study applied a policy–utility machine learning framework to evaluate mesh size regulations for the Marbled Flounder trap fishery in southern Republic of Korea. By integrating CPUE, the proportion of immature fish, and the bycatch ratio into a unified policy–utility function, the approach transcended traditional length-based selectivity analysis by explicitly quantifying trade-offs between productivity and conservation objectives. The gradient boosting model, which was trained on 2016 experimental data and externally validated with 2015 data, demonstrated strong predictive accuracy. Thus, the ΔU structure reliably captures the combined effects of mesh size, seasonal conditions, and catch composition.

The empirical optimization of policy weights emphasized stronger penalties with a higher catch of immature fish, and variable importance analysis identified the immature proportion as the dominant utility driver. Simulated scenarios showed that the optimal mesh size shifted from 65 mm in 2015 to 80 mm in 2016, which reflected interannual changes in size structure and bycatch. However, policy regret analysis indicated that the utility loss from choosing 80 mm mesh over 65 mm mesh was minimal (ΔU ≈ 0.12–0.15). The relative regret remained below the robustness thresholds in all scenarios and supported 80 mm as a stable, low-risk regulatory option.

Conceptually, the proposed utility–regret framework complements higher-level EAFM tools, such as GFMI, SAFE, and ERAEF, by operating at the within-gear design scale. This allows the robust selection of mesh size rules under uncertainty. Although this framework is limited by single-area, short-term data and the interpretability constraints of machine learning models, it provides a practical template for expanding trap selectivity research toward operational, EAFM-aligned mesh size regulation.