Competence Estimation Impact in Multi-Agent Systems

Calibrated trust is the human-machine teaming concept that machine-learned model predictions should not be over or under trusted (McDermott, 2018). Application of machine-learned model predictions is often assessed through model reported confidence or other model specific confidence metrics, with the assumption being applied that model confidence is calibrated (Guo, 2017). However, model confidence estimates do not account for uncertainties due to training data distribution, class representation, and model design which impact prediction suitability (Rajendran & LeVine, 2019).

It is not adequate to apply a model that performs well during design and development without considering its suitability during deployment. Differences in data used in model development and evaluation from real world inputs or deployment “in the wild” can affect accuracy in prediction (Koh, 2021; Amodei, 2016). To that end, the WILDS benchmark datasets capture distributional shifts that may occur while models are deployed (Koh, 2021). These datasets were curated to explore the development of models that may be more robust to distributional shift over time (Koh, 2021). The differences may arise due to natural distribution shifts over time, such as differences in imaging due to instrumentation, data collection procedures, or environmental conditions (Koh, 2021). The impacts of models not robust to distributional change can lead to problematic errors such as unfair, biased, offensive, or harmful predictions (Koh, 2021; Amodei, 2016). Additionally, similar challenges arise from transfer learning, where models may be trained on simulation or other alternative data sources that do not capture real world data (Amodei, 2016).

Consideration of distributional shifts and applicability of training data to real world data in model deployment motivate an approach for providing a user or downstream application indication of where the training data and inputs “in the wild” are misaligned. Point-wise competence estimation methods enable this indication of model prediction suitability to be provided at the time of decision. During test and evaluation, point-wise competence estimates provide an analytical indication of operating conditions where model predictions are expected to be reliable based on the model’s training data. Furthermore, point-wise competence estimates may be provided during deployment to calibrate trust in model predictions.

Model competence estimation extends the current application of model reported confidence to ensure that models are trusted on inputs for which they have appropriate prediction capability and corresponding training data. Point-wise competence estimation methods have enabled integration of competence estimates into human-machine teaming applications to ensure appropriate application of a model for a given input. Point-wise competence estimates are normalized to the interval [0, 1], where larger values indicate a greater estimated suitability of the underlying model for the corresponding input.

As a noteworthy example, the Accurate Layerwise Interpretable Competence Estimation (ALICE) point-wise competence accounts for distributional, data, and model uncertainties for classification models (Rajendran & LeVine, 2019). The approach produces an ALICE score for each input to a classification model that is formed from three factors including an in-distribution comparison to training data. If the value of the score exceeds a user set threshold, the model is deemed competent for that specific input.

In point-wise competence estimation, each input point x is compared to the distribution of data used to train a machine-learned model. Let the variable D be the event that x is in distribution to the training data of specified model. Then p(D|x) is the probability of the event that x is in the training data distribution. For a point-wise competence estimation technique, this probability is referred to as the in distribution probability and is empirically derived. Other factors are incorporated tied to predictive capability for a model.

A point-wise competence estimation approach leveraging an in distribution and prediction interval mask was developed for regression models and described in detail in McFadden (2026). The method may be applied across all phases of development and deployment to calibrate trust in use of model predictions. The competence estimates produced from the approach have exhibited a consistent inverse relationship with absolute error across multiple publicly available datasets, as shown in previous publications (McFadden, 2026). Consequently, higher competence estimates are expected to correspond to more reliable predictions.

The work in this document uses previously developed point-wise competence estimation methods and focuses on their integration into multi-agent systems.

In the cases where an individual model is not deemed competent, then a human decision, algorithm, heuristic, or policy should be integrated as an intervention. Alternatively, a different machine-learned model may be applied to the input, which would be expected to be more suitable for prediction. Ensemble learning approaches offer a complementary application of competence because multiple models are available for use.

Ensemble learning methods employ the predictions of multiple base models to produce a single prediction (Kundu, 2024). The predictions of the base models are combined, blended, or selected to produce a single prediction from the ensemble. Ensemble learning predictions benefit when the ensemble is deliberately formed from diverse base models such that the models will have different decision boundaries (Polikar, 2006).

In ensemble classification, each base model produces a prediction label. Often, there is consensus among the models in an ensemble, e.g., they all agree on the predicted class of an image. When there is non-consensus among the base models, a voting scheme is applied. Base classification models are expected to produce a confidence with the prediction. In non-consensus, the prediction with the highest confidence may be selected.

In previous work, the ensemble learning approach was extended to incorporate individual base model competence within the fusion of base model predictions (McFadden, 2025). The resulting ensemble classification approach demonstrated that integrating point-wise competence into the voting scheme improved performance compared to the highest confidence selection strategy (McFadden, 2025).

Subsequent work extended the point-wise competence integration to ensemble regression, where higher competence estimates receive greater weight in an ensemble (McFadden, 2026). During model deployment, point-wise competence estimates are used to weight ensemble predictions. Suppose there are T ≥ 2 base models. For an input point x, each model j = 1, …, T produces a prediction yj and an associated competence estimate S_j∈0, 1. The competence-weighted ensemble prediction is given by,

The competence-enhanced ensemble regression approach demonstrated reduced prediction errors compared to an equal-weight ensemble, in which predictions from base models are averaged equally (McFadden, 2026).

In the present work, these traditional ensemble fusion strategies serve as background; the following approach and experiment evaluate the research question of how competence informed coordination compares to a static specialization-based coordination strategy in multi-agent systems. In multi-agent systems, a coordinating agent applies the competence-weighted combination principle when integrating predictions from multiple agents.

The rapid adoption of multi-agent systems has motivated application of architectures where many agents exhibit interdependencies while individually applying machine-learned models and tools to complete a task (Sapkota, 2026). Coordinating agents can break down and assign tasks among multiple other agents leading to orchestration and collaboration within the architecture. Individual agents may have roles such as data gathering and cleaning, modeling and prediction, and generating findings with reasoning.

Agents may have specialties in narrow and well-defined tasks (Sapkota, 2026). An example architecture for a self-driving car application is shown in Figure 1 where multiple agents contribute route planning, obstacle avoidance, and vehicle status. As depicted in the example, regression and classification models provide predictions to agents. A need remains to ensure model predictions are applied appropriately as machine reasoning is used to consider information from multiple text sources and predictive models. Hallucination, misinterpretation, and over-trust in results can lead to false reasoning and conclusions.

Multi-agent systems incorporate predictions and results from multiple agents that negotiate or compete to accomplish a task. Similar to ensemble learning, the outputs from multiple agents may be combined and blended for decision-making. In that sense, a multi-agent system can be viewed as an ensemble from many agents reliant on specialized models. An agent’s specialized models may not be suitable for prediction in the specific conditions associated with a query, motivating the need to estimate and incorporate point-wise competence-based-weighting when selecting and combining agent outputs.

Figure 1: Example self-driving car multi-agent system architecture.

The following approach and application were motivated by the research question of whether incorporation of point-wise competence estimation could improve selection, coordination, and blending of agent responses in multi-agent systems compared to a baseline static specialization-based expertise assignment.

Multi-agent systems employ agent expertise by assigning tasks based on each agent’s capabilities and the requirements of the task. While division of tasks by specialization enables capability-based assignment, it does not fully account for model, data, and distributional uncertainties that are accounted for through point-wise competence estimation. The purpose of the described concept is to enhance reasoning, collaboration, and coordination in multi-agent systems through consideration of point-wise competence estimation for dynamic assessment based on current conditions and inputs.

As noted previously, agents may employ classification or regression model predictions in their reasoning processes. In the proposed approach, models report both prediction and an associated point-wise competence estimate to agents for each input. Agents must incorporate their competence estimates alongside their predictions when generating and communicating their results. The reasoning process includes gathering contextual information such as holiday schedules, producing model predictions, and incorporating point-wise competence estimates into coordination decisions. Competence weighting therefore becomes a component of the broader reasoning process.

In the proposed architecture, coordination refers to the process of selecting and combining agent responses. Coordinating agents use reported point-wise competence estimates to weight, select, or reroute among agent responses. Depending on the needs of the task, the coordination policy may weight predictions from multiple agent responses, select the response associated with the highest competence, or reroute when competence estimates are below a specified threshold, indicating that an agent’s underlying model is not suitable for the current input. The specific coordination policy, including any competence thresholds, is application dependent and is therefore not prescribed by the proposed architecture.

By incorporating point-wise competence estimation into agent interactions, this approach enables multi-agent systems to dynamically assign expertise based on the estimated competence of each agent’s underlying model for the current conditions and inputs. Consequently, selection and combination of agent responses can account for model training data coverage, out of distribution comparison, and model feature selection. Competence estimate integration empowers the system with an internal mechanism to dynamically evaluate if an underlying prediction being used by an agent is suitable for a particular combination of conditions associated with a system query or user prompt.

To demonstrate the approach, a dataset of bike sharing system trends (Fanaee, 2014) was used. Bike sharing programs offer rentals which may be returned at any of multiple sites around a city, not necessarily the site the bike was acquired from.

The version of the dataset hosted by the University of California, Irvine (UCI) Machine Learning Repository was used. The dataset includes two years of logged reservations within the Washington, D.C. Capital Bike Sharing system. The dataset includes 17,379 hourly observations, where counts of registered commuters with a membership subscription and casual riders are recorded. The dataset includes season, month, weekday versus weekend, holiday impact, hour, and multiple weather conditions. There were 22 data points for which the humidity field was not populated which were removed and not considered in training or test. A sample size of 17,357 remained for training and test. Additionally, two weather-related-fields in the dataset were scaled as stored in the originating repository. To facilitate interpretation of the analysis results and corresponding figures, two weather features were converted to their original units. Specifically, temperature was converted to degrees Fahrenheit and wind speed was converted to kilometers per hour. The scaled values were used in model training and test.

Table 1: Absolute Pearson correlation with bike sharing rider counts, from 17,357 points.

Table 1 shows the absolute Pearson correlation coefficients of the bike sharing dataset features with bike sharing system rental counts. Three rider count values are supplied in the dataset; the overall rider count is the summation of the registered riders and casual riders. As expected, the linear relationship between the overall count with the additional two count features is strong. Count values in Table 1 were included only for exploratory correlation analysis and were not used as predictor variables in machine-learned models. The weather fields such as temperature and humidity exhibit a moderate linear relationship with overall rider counts. The hour of day shows a moderate linear relationship with rider counts as well. Among the weather-related features, wind speed exhibits the weakest linear relationship with rider counts.

Figure 2 shows the trends in both registered and casual rental counts by hour of the day and temperature in Fahrenheit. For the purpose of illustration, temperatures are shown in two categories. Points in red indicate temperature above 60 degrees Fahrenheit while points in gray indicate temperature at or below 60 degrees Fahrenheit. As indicated by Table 1, bike rider count is moderately correlated with both weather and time of day. In Figure 2, the registered rider counts show a pattern of increased usage during peak commuting times (morning and late afternoon) not observed in casual rider counts.

Figure 2: Registered (left) and casual (right) bike rider counts by hour and temperature.

The weather-related features include sufficient detail to account for variation beyond broader season, including wind speed, humidity, temperature, and a categorical field of weather situation. Specifically, the weather situation data provides contextual weather information such as clear skies, cloudy skies, or precipitation. Models suitable for each weather situation may be formed; however, models tailored to a specific weather situation, e.g., heavy rain, are not expected to generalize well to the broader dataset.

Although the dataset captures a variety of factors that affect bike sharing system rider counts, additional value could be gained from the inclusion of specific bike sharing system pick up and drop off locations within Washington, D.C., as well as broader geographic variation beyond the city limits and the surrounding regions.

The bike sharing system is used to demonstrate the proposed approach for integration of point-wise competence estimation in a multi-agent system. An example multi-agent systems architecture intended to respond to queries on the bike sharing system is shown in Figure 3. Multiple agents may be responsible for gathering contextual information such as weather forecasts, holiday schedules, and special event calendars. This information is provided to other agents and incorporated into the models they use in reasoning on predicted demand. In response, the models reply with predictions and point-wise competence estimates. Point-wise competence integration empowers the system with an internal mechanism to analytically evaluate if a model used by an agent is competent for a particular combination of conditions, e.g., weather, hour, holiday. The information is used for weighting agent responses and dynamically coordinating expertise based on specific conditions applicable for a query, e.g., humidity.

Figure 3: Bike sharing system multi-agent system architecture.

To that end, the bike sharing system application motivates the use of specialized agents based on tailoring of weather, temporal, and additional demand related factors, such as holiday schedules. Each agent may employ classification and regression models using a subset of the overall features available and subsets of the training data relevant to an assigned expertise. The categorical feature ‘weather situation’ was chosen as an example specialization for bike sharing system demand prediction.

Four weather situations are represented including clear skies, mist or cloudy, light precipitation, and heavy precipitation. In the dataset documentation, the distinction between light and heavy precipitation is qualitative rather than based on a quantitative threshold. Light precipitation is described as “Light Snow, Light Rain + Thunderstorm + Scattered clouds, Light Rain + Scattered clouds” whereas heavy precipitation is described as “Heavy Rain + Ice Pallets + Thunderstorm + Mist, Snow + Fog” (Fanaee, 2014). Figure 4 shows the counts by weather situation from 17,357 points in the bike sharing dataset. As few data points were available for heavy precipitation, a model specific to the corresponding weather situation was not used.

Figure 4: Bike sharing dataset counts by weather situation.

A baseline static specialization-based coordination strategy assigns each query to the agent or model corresponding to its predefined area of expertise, without considering point-wise competence estimates. In the bike sharing dataset demonstration, the baseline coordination strategy assigns each query to the agent using the weather situation model corresponding to the query weather situation. In contrast, the proposed approach dynamically combines predictions from all candidate weather situation models using their point-wise competence estimates. The following study is used to show the trades in a contextual, specialization based multi-agent system assignment and coordination approach compared to the proposed approach for point-wise competence estimation.

An example query is considered on the expected bike sharing system demand in Washington, D.C. for casual bike riders during a winter holiday weekend with clear skies. As the predicted target, casual rider counts were selected for study as they showed the strongest overall linear relationship to the weather related variables. In a standard specialization based agent assignment, predictions would be assigned to the agent specialized to the query weather situation, e.g., the agent using the clear skies model.

Although the proposed architecture accommodates both classification and regression models, the bike sharing demonstration uses regression models to predict casual rider counts. Three random forest regression models were formed using MATLAB TreeBagger, each tailored to a different weather situation. The weather situation models did not have weather situation as a feature in design, and instead incorporated the features associated with weather such as humidity, temperature, season, and hour.

A test subset of the bike sharing system data is used to verify the trends in competence estimates and the response to the query under consideration. The bike sharing dataset was randomly partitioned into equal sized training and test sets, where the training partition was reserved for model formation and the test partition was withheld for evaluation. Figure 5 shows the rider counts over time increased during the second year, indicating a more mature operating period for both the train and test partitions. Consequently, only the 4,381 second year observations within the training partition were used for model formation because they reflected higher and more stable casual rider counts, reducing the influence of early registered rider adoption on the demonstration. All 8,678 observations in the test partition, including both years, were retained for evaluation.

Figure 5: Train (left) and test (right) total and casual rider counts over time.

The example query being considered pertains to winter holiday weekends with clear skies. A model primarily trained on light precipitation weather situation data should produce low competence estimates for a clear skies day. Figure 6 shows the point-wise competence estimates of each weather situation model for the 44 test points corresponding to winter holiday weekends with clear skies. Since the query conditions are more consistent with the training data used to form the clear skies situation model than with the ‘light rain/snow’ model training data, the clear skies situation model produces higher competence values. The January 2011 test point shows a low competence value because the combination of hour and temperature is rare among the data used to train the ‘clear’ model.

Figure 6: Point-wise competence estimates of each weather situation model for 44 test points corresponding to winter holiday weekends with clear skies.

Figure 7 shows the distributions of the point-wise competence values of each weather situation model for the 44 test points corresponding to winter holiday weekends with clear skies from both years of the bike sharing dataset. The blue rectangles indicate the interquartile range of each distribution, the red horizontal lines indicate the median of each distribution, the black whiskers extend to observations within 1.5 times the interquartile range, and the plus symbols indicate outliers. As expected, the ‘clear’ model generally produces higher point-wise competence values than the ‘mist/cloudy’ and ‘light rain/snow’ situation models. The interquartile range of the ‘clear’ model lies entirely above the interquartile range of the ‘light rain/snow’ model, indicating a pronounced separation in the central values of their competence distributions. The interquartile ranges of the ‘clear’ and ‘mist/cloudy’ situation models exhibit partial overlap, reflecting that either model may be suitable for some query observations. The ‘light rain/snow’ model has the lowest median value and a narrow interquartile near zero, along with a limited number of higher outlier observations. The ‘light rain/snow’ model shows substantially lower competence values than the other weather situation models consistent with the query conditions being less representative of the training data used to form the ‘light rain/snow’ model.

Figure 7: Distributions of the point-wise competence estimates of each weather situation model for the 44 test points corresponding to winter holiday weekends with clear skies.

The original query concerns the projected casual rider counts during a winter holiday weekend with clear skies. The impact of the proposed approach on the query response is evaluated through comparison to the baseline coordination strategy, which uses the ‘clear’ weather situation model predictions. Figure 8 shows three thin solid lines colored by weather situation model, the test point target values as a bold violet line, the competence informed ensemble prediction as a dashed green line, and the equal weight ensemble as a dash-dot light blue line. The predictions from both ensemble methods are rounded to whole numbers. The ‘clear’ model (thin blue line), as the baseline coordination strategy, better aligns with casual rider count than the two other individual weather situation models. However, the clear skies situation model over predicts casual rider counts at multiple test points. Point-wise competence integration better aligns to the observed target values of the test points. Consistent with prior work (McFadden, 2026), the incorporation of competence to weight the weather situation models aligns better with the test point target values than an equal weighted blending of the weather model predictions. The point-wise competence enhanced ensemble prediction line produces the most consistent prediction with the test point target values. Moreover, at the test point where the three weather situation models have point-wise competence under 0.2, the predictive performance of the point-wise competence ensemble prediction is best aligned to the target observation. To show the weighting inputs, Table 2 lists the weather situation predictions and competence low competence test point, and the resulting ensemble predictions.

Table 2: Model predictions at the low (<0.2) competence test point. The point-wise competence values are in parentheses.

Figure 8: Prediction of casual rider counts based on the weather situation models and their ensemble prediction in comparison to the query 44 test point target values.

Mean average error (MAE) and root mean squared error (RMSE) are standard metrics used to evaluate predictive performance. Since the two metrics capture errors differently, they provide complementary measures of model accuracy. The ‘clear’ weather situation model (the baseline approach) and the competence enhanced ensemble prediction were both compared to the casual rider count observations for holiday weekends with clear skies. Table 3 shows the MAE and RMSE of the ‘clear’ weather situation model compared to the integration of point-wise competence estimation by year. In both years, point-wise competence-informed coordination consistently outperformed both static specialist assignment and equal-weight ensemble averaging. The largest gains occurred in Year 1, where the competence estimates compensated for distributional differences between the training and test data. In the first year, the performance improvement for the prediction of casual rider counts on the holiday weekend with clear skies test points is considerable, with improvement to MAE of 29.9% and an improvement to RMSE of 30.4%. Although improvements were smaller in Year 2, competence-informed coordination still produced lower MAE and RMSE, demonstrating that dynamic competence weighting provides benefits beyond the baseline specialist selection. The point-wise competence estimates contribute to dynamic assignment of agent expertise and improved prediction accuracy.

By incorporating point-wise competence estimation into the reasoning process, predictions from multiple specialized models were weighted according to their estimated suitability for the specific inputs. Beyond the baseline specialty assignment, point-wise competence improved prediction response to the query. The benefits are in both blending of results and coordination. A coordinating agent should not necessarily select the agent with the most obvious specialization as may be done in a baseline static specialization-based coordination strategy where pre-defined agent and model expertise is used.

Table 3: Comparison of model predictive performance on the test points of casual rider counts on holiday weekends with clear skies by year.

Point-wise competence estimation was initially developed for consideration in human-machine teaming and autonomous systems to calibrate trust in the application of machine-learned model predictions. Early work explored computational trades in implementation permitting point-wise competence to be available at the time of decision for both classification and regression models. Subsequent research demonstrated improved performance when point-wise competence estimation is incorporated in the fusion of predictions from multiple, diverse models in ensemble learning. The approach and application results suggested that point-wise competence estimation may have a broader role and impact when used within multi-agent systems.

The exploration presented in this work was motivated by the research question of whether incorporation of point-wise competence estimation could improve selection, coordination, and blending of agent reasoning in multi-agent systems compared to a static specialization-based expertise assignment. An approach for integration of point-wise competence into multi-agent systems was described and illustrated. Point-wise competence estimation provides a dynamic, input-dependent indicator of model suitability that can be incorporated into agent reasoning and coordination decisions.

In the bike sharing demonstration, competence estimation integration was applied to capture combinations of conditions for which a model may be suitable to predict results, including comparison to training data and feature selection impacts for an input. The results indicate that point-wise competence estimation integration can improve both the combination of agent responses and predictive accuracy of the resulting coordinated prediction. The demonstrated competence integration approach exhibited improved predictive performance compared to the implemented baseline coordination strategy in which queries were assigned to the weather situation model corresponding to the query weather situation, e.g., the clear skies model. Furthermore, point-wise competence estimation integration improved accuracy in prediction compared to a strategy assigning responsibility to a single specialized model, enabling the system to better capture nuanced trends in casual rider counts. In the demonstration, both MAE and RMSE were reduced, indicating improvement in predictive performance. The results of the bike sharing demonstration indicate that point-wise competence estimation integration can improve both the combination of agent responses and predictive accuracy by enabling coordination to adapt dynamically to the conditions associated with a specific query.

The weather situation example represents one of many specialized predictive models which could be developed from the bike sharing dataset or other datasets. For instance, additional insight and further tailoring of predictions could be achieved from the inclusion of bike sharing system pick up and drop off locations within Washington, D.C., as well as broader geographic variation beyond the city limits and the surrounding regions. The additional data would enable location and region specific information accounting for tourism and seasonal demands as well, while a coordinating agent dynamically weights predictions based on estimated competence.

Processing was performed in MATLAB on a 13th-generation Intel Core i7-13700P (base frequency 1.90 GHz), 32 GB RAM, x64-based system.

MATLAB R2024b was used to process data, create base models, prototype competence scores, and generate figures for the results shown. The default algorithm and seed for the MATLAB built in random number generator were applied for reproducibility.

Charts corresponding to this article were presented at the Defense and Aerospace Test and Analysis Workshop (DATAWorks) conference in Alexandria, VA on April 22, 2026.

The University of California, Irvine (UCI) Machine Learning Repository hosted the Bike Sharing dataset used for illustration of this concept.

1. Amodei, D. and C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Man´e, “Concrete Problems in AI Safety,” arXiv preprint arXiv:1606.06565, 2016, accessed March 1, 2026. [Online]. Available: https://arxiv.org/abs/1606.06565.
2. Fanaee, T. and J. Gama, “Event labeling combining ensemble detectors and background knowledge,” Progress in Artificial Intelligence, pp. 1–15, 2013.
3. Guo, C. and G. Pleiss, Y. Sun, and K. Q. Weinberger, “On Calibration of Modern Neural Networks,” arXiv preprint arXiv:1706.04599, 2017, accessed February 6, 2026. [Online]. Available: https://arxiv.org/abs/1706.04599
4. Koh P. W. and S. Sagawa, H. Marklund, S. M. Xie, M. Zhang, A. Balsubramani, W. Hu, M. Yasunaga, R. L. Phillips, I. Gao, T. Lee, E. David, I. Stavness, W. Guo, B. Earnshaw, I. Haque, S. M. Beery, J. Leskovec, A. Kundaje, E. Pierson, S. Levine, C. Finn, and P. Liang, “WILDS: A Benchmark of in-the-Wild Distribution Shifts,” in Proceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol.139. PMLR, 18–24 Jul 2021, pp. 5637–5664, accessed February 14, 2026. [Online]. Available: https://proceedings.mlr.press/v139/koh21a.html
5. Kundu, R. “The Essential Guide to Ensemble Learning,” V7 Labs, 11 Jan 2024, Retrieved from: https://www.v7labs.com/blog/ensemble-learning-guide.
6. McDermott P. and C. Dominguez, N. Kasdaglis, M. Ryan, I. Trahan, and A. Nelson, “Human-machine teaming systems engineering guide,” The MITRE Corporation, Bedford, MA, Tech. Rep., 2018, accessed February 6, 2026. [Online]. Available: https://www.mitre.org/sites/default/files/2021-11/prs-17-4208-human-machine-teaming-systems-engineering-guide.pdf
7. McFadden, F. R., “Competence Measure Enhanced Ensemble Learning Voting Schemes,” The ITEA Journal of Test and Evaluation, vol. 46, no. 3, Sep. 2025, accessed February 6, 2026. [Online]. Available: https://itea.org/journals/volume-46-3/competence-measure-enhancedensemble-learning
8. McFadden, F. R., “Competence Estimation Enhanced Ensemble Learning with Base Regression Models,” 2026, available at: https://dx.doi.org/10.2139/ssrn.6962188
9. Polikar, E. “Ensemble based systems in decision making,” in IEEE Circuits and Systems Magazine, Vol. 6, No. 3, 2006.
10. Rajendran, V. and W. LeVine. “Accurate Layerwise Interpretable Competence Estimation,” In Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alché-Buc, E. Fox, R. Garnett, editors, Vol. 32, Curran Associates, Inc., 2019.
11. Sapkota R. and K. I. Roumeliotis, and M. Karkee, “AI Agents vs. Agentic AI: A Conceptual Taxonomy, applications and challenges,” Information Fusion, vol. 126, p. 103599, Feb. 2026, accessed February 6, 2026. [Online]. Available: http://dx.doi.org/10.1016/j.inffus.2025.103599