acceptodds
Under review as a conference paper at ICLR 2027

DYNABEE: : Dynamical Systems for Modeling Honey Bee Colony Behavior

Abstract

Given sparse and noisy measurements, how can we model, forecast, understand the colony behavior? Domain experts strictly demand that the method is explainable, fast, and accurate. We propose DYNABEE, which consists of a carefully designed dynamical system of differential equations, along with their stability analysis. We developed the equations from first principles, in close collaboration with entomologists. As a result, DYNABEE has only few, perfectly interpretable parameters to fit (foraging rate, mortality rate, etc). Thus, DYNABEE has the following desirable properties: (a) Accurate: experiments on the largest available dataset from real hives show good accuracy; (b) Explainable: our model parameters have physical meaning; (c) Scalable: DYNABEE is linear on the input size; (d) Transferable: the fitted parameter values give good forecasting for new bee colonies, too.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Honeybees, Dynamical System, Forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8U8HOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a model called DYNABEE for forecasting the behaviour of honey bee colonies. Accurate forecasting is important to beekeepers, as it will inform them on when to treat bee colonies and improve their health. Previous approaches have been proposed, but this paper shows that its approach is better on the following properties that make the forecasting model useful: (a) Accuracy, (b) Explainability, ( c), Scalability, and (d), Transferability. Specifically, their approach develops the model using systems of equations, instead of using black-box approaches (deep learning).

Critical strengths

The authors suggest that the proposed solution, DYNABEE, can do tasks (forecasting in this case) that black-box (deep-learning methods) cannot do. They go and compare their approach against those that use black-box approaches and show how theirs is better. This is significant.

Critical weaknesses

None.

Questions for authors

None.

Reviewer_h9vyOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of modeling, forecasting, and understanding honeybee colony behavior using sparse and noisy hive inspection records. The authors propose DYNABEE, a biologically motivated dynamical system based on ordinary differential equations (ODEs) that model honey reserves, egg laying, larval development, adult bee populations, mite infestation, and treatment effects.

The main contributions are: (1) an interpretable mathematical model with parameters representing biological processes, (2) equilibrium and stability analysis to study colony dynamics, (3) the ability to investigate hypothetical scenarios by modifying biologically meaningful parameters, and (4) a computationally efficient forecasting approach that can potentially operate with limited observations.

For validation, the authors use two real-world datasets: six hives for estimating shared model parameters and eleven additional hives to evaluate transferability without retraining those parameters. They compare one-step-ahead forecasting accuracy against traditional forecasting baselines, DeepAR, and variants of their proposed model using RMSE.

The experimental results suggest promising forecasting performance and some degree of transferability across colonies. However, the validation is based on a relatively small number of hives, and the biological accuracy of the predicted equilibria and hypothetical intervention scenarios is not independently established.

Critical strengths

1. Biologically motivated and interpretable modeling approach.
The paper proposes a relatively simple ODE-based model that incorporates meaningful biological processes, including egg laying, foraging, honey consumption, larval development, mortality, and mite treatment. I consider this a valuable contribution because the model offers more than forecasting: its parameters and equilibrium analysis can potentially help researchers and practitioners understand colony dynamics and investigate hypothetical scenarios. Although the biological validity of these scenarios requires further validation, the interpretability of the approach is a meaningful strength.

2. Real-world data and transferability across colonies.
The authors evaluate their model using real hive inspection data, which are difficult and time-consuming to collect. I appreciate that the authors have also made their data and implementation publicly available. Furthermore, they estimate the shared model parameters using six hives and apply them to eleven additional hives without retraining those parameters. Although the experimental scale is limited and transferability is demonstrated under relatively constrained conditions, the results suggest that the model could be useful in practical settings where only limited historical observations are available.

Critical weaknesses

1. Insufficient methodological clarity and experimental validation.

My main concern is that the manuscript does not provide sufficient information to independently assess the reliability of its results. The authors report 95% confidence intervals without explaining how they were computed. Furthermore, the data collection methodology is not adequately described, particularly how hive measurements were obtained and why some observations have up to four decimal places of precision. The paper also relies heavily on graphical comparisons and qualitative claims such as "good accuracy" without providing sufficiently detailed numerical results. These limitations make it difficult to evaluate the reported forecasting performance and its statistical reliability.

2. Mathematical inconsistencies and insufficient justification of the model's biological validity.

The paper emphasizes mathematical interpretability and stability analysis as central contributions. However, several mathematical statements appear inconsistent with the proposed equations, particularly the equilibrium expressions and stability claims in Lemmas 2–4. Additionally, some estimated parameters fall outside their stated biological ranges without adequate explanation. The what-if scenarios are also not independently validated against observed biological interventions. Since mathematical and biological interpretability are among the paper's main contributions, these issues raise concerns about the validity of its conclusions and substantially influence my recommendation toward rejection.

Questions for authors

1. Mathematical correctness and stability analysis

Several mathematical results appear inconsistent with the proposed equations. In particular, Lemma 4 seems to omit the starvation throttling factor e(H∗)e(H^*) from the equilibrium expression for healthy adult bees. Lemma 2 appears to characterize points as saddle equilibria that are not necessarily equilibria, and Lemma 3 seems to require additional assumptions for stability. Furthermore, the numerical equilibrium reported for Lemma 1 does not appear to match the provided parameters. Could the authors clarify these discrepancies and demonstrate that the reported analytical results, numerical calculations, and implementation are consistent? Do any corrections affect the conclusions?

2. Computation of the 95% confidence intervals

How were the reported 95% confidence intervals in Figure 1 calculated? Do they represent uncertainty in the estimated parameters, model predictions, or something else? Were these intervals evaluated for empirical coverage on held-out observations? This is important for assessing the statistical reliability of the reported forecasting results.

3. Data collection methodology and measurement precision

Could the authors explain how the hive inspection measurements were collected and converted into numerical values? I examined the released dataset and noticed that some bee-frame measurements contain up to four decimal places. How was this level of precision obtained from manual inspections? Are these direct observations, estimates, interpolated measurements, or processed values? I would also appreciate clarification on how mite infestation measurements were converted into the model's representation of mite-infected bee frames.

4. Quantitative forecasting results and transferability

Could the authors provide exact numerical forecasting errors, preferably with variability across individual hives, rather than relying primarily on graphical comparisons? Additionally, the transferability experiment uses parameters estimated from six hives and applies them to eleven additional hives. Were these datasets collected from the same geographical location or under similar environmental conditions? How much target-hive observation data is required for initialization, and is the forecast horizon one day or the interval until the next inspection? These details would help determine the extent to which the results demonstrate generalizability.

5. Biological validity of the estimated parameters and hypothetical scenarios

The authors emphasize that the model parameters have biological interpretations and can be checked against expected ranges. However, Table 2 reports values for c7′c'_7 and c8′c'_8 that fall outside their stated ranges. Could the authors explain these discrepancies and whether the imposed parameter relationships are biologically justified? Additionally, is there any independent evidence supporting the quantitative outcomes of the what-if scenarios, particularly changes in larval populations resulting from different egg-laying rates?

These are clarification requests concerning the manuscript's central claims rather than suggestions for future work. The responses, particularly regarding mathematical correctness, uncertainty estimation, and data collection, could materially affect my final assessment.

1 comment

Q1. (a) We updated Lemma 4 to match the revised equation set (7) - (14). Please refer to Section 3.3.1.

(b) Please refer to Appendix A.3 for the proof of Lemma 2

(c)Please refer to Appendix A.4 for the proof of Lemma 3

(d)We thank the reviewer for this observation, and we have addressed the comment. Please refer to Section 3.1.1 in the revised manuscript.

(e) None of the corrections affect the conclusions.

Q2. The shaded band in Figure 1 in our original submission was computed by x^±1.96σ/n\hat{x} \pm 1.96 \sigma / \sqrt{n} where x^\hat{x} is the forecasted point, σ\sigma is the standard deviation of the reconstruction errors, and nn is the window size. It reflects the uncertainty of the mean reconstruction error over the window. Therefore, it is not expected to contain 95% of new observations, and it should have been described explicitly as a confidence interval for the mean reconstruction error. In the revision, we replace this band with x^±σ\hat{x} \pm \sigma which reflects the 68% prediction interval. We also report the empirical coverage in Appendix A.8. In summary, 50-100% of our forecasts fall in this band (Table A10).

Q3. (a) Colony inspections were conducted approximately every two to three weeks. For each hive, the beekeeper records the number of frames covered by honey, larvae (brood), and adult bees, as integer frame counts, together with a mite measurement (alcohol wash, as a percentage of bees sampled).

(b) Values with decimals (up to four places) are linearly interpolated between consecutive inspections to obtain a regular time grid. They are not direct measurements.

(c) Mite infection intensities were quantified using the alcohol wash method, which is widely used by beekeepers to monitor and manage Varroa infestations. Briefly, we collected a subsample of approximately 400 adult worker bees from each colony and counted the number of V. destructor mites recovered from these bees. This provided an estimate of mite infection intensities, expressed as the average number of mites per sampled bee. DynaBee models the honey bee colony at the population level, treating it as a superorganism rather than tracking individual bees. Accordingly, the model variable (B_M) represents the population of mite-infested adult bees, expressed in equivalent bee frames. To convert the empirical measurements into mite intensities, we assumed that each recovered mite corresponded to one infested adult worker. We then multiplied the observed mite-to-bee ratio by the total number of frames occupied by adult bees (B): Mite infested bee frame B_M = Total bee frame B × (no of mites / bees sampled).

We added this explanation as a footnote in Section 5. Q4. (a) We have added Tables A2-A9 (appendix A.7) reporting RMSE for each method on every individual hive, for both Dataset 1 and Dataset 2.

(b) Yes, both datasets were collected from the same geographical location and under similar environmental conditions. We added this explanation in Section 5.

(c) For initialization, we use the inspection data from six hives in the first dataset. For each hive, the first four recorded inspection points, together with eight linearly interpolated points for H, L, B_M , and B_H , are used for parameter estimation.

The forecast horizon is the next inspection point, corresponding to a prediction horizon of approximately two to three weeks, which matches the scheduled hive inspection interval Q5. (a) The ranges listed for (c'_7) and (c'_8) in Table 2 were transcription errors. We updated the table.

(b) We do not have independent empirical evidence (e.g., controlled experiments varying queen egg-laying rate) that validates the quantitative outcomes of the what-if scenarios, and we do not claim that they are predictive of real colonies. The scenarios are intended as simulation, model-based explorations: they show how the model’s state variables respond when c4 is varied, so that domain experts can judge whether the direction and relative size of the response is biologically plausible.

Reviewer_UbWfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper iis about DynaBee, a mechanistic approach for forecasting honeybee colony dynamics using systems of ordinary differential equations. The central motivation is to provide accurate and biologically interpretable forecasts from sparse, noisy, manually collected hive inspection records, rather than relying on expensive sensor installations or black-box forecasting methods. The authors provide equilibrium and stability analyses and propose an optimization algorithm that estimates shared biological parameters and hive-specific initial conditions through nonlinear least squares. The paper addresses a meaningful problem and presents an appealing application of interpretable dynamical systems. However, important issues concerning the mathematical formulation, theoretical claims, biological validity, and experimental evaluation need to be resolved.

Critical strengths

The paper targets a real-world forecasting problem where observations are sparse, noisy, and expensive to collect.

A major advantage is that the model parameters have explicit biological interpretations. This makes it possible to inspect the fitted dynamics, simulate interventions, and analyze equilibria. This is particularly valuable in applications where domain experts need to understand predictions before acting.

The dynamical-systems perspective allows investigation of long-term population behavior and biologically motivated what-if scenarios. These capabilities could be valuable for understanding colony dynamics, provided that the underlying mechanistic equations and fitted parameters are validated.

Critical weaknesses

1/ The full model appears to omit the loss of larvae when they develop into adult bees. Am i mistaken ? and why ? As written, my understanding is that larvae contribute to the creation of adult bees without leaving the larval compartment. Unless the authors intend a different biological interpretation of these variables, this creates a substantial inconsistency in the population flows.

2/ I do not understand the proofs of Lemma 2 and Lemma 4 (where is e(H^) in the expression of B^_H ?

3/ the eigenvalues \lambda_\inf is of sign c_8-c_9 so it is not always positive ...

4/ The paper could benefit from per-hive results, uncertainty intervals on the comparative performance, and statistical tests accounting for the dependence between repeated measurements from the same hive

Questions for authors

1/ Why do Equations (9)–(10) not contain negative emergence terms corresponding to the adult population gains in Equations (11)–(12)? Is this intentional, and how is the biological flow from larvae to adults conserved?

2/Can the authors revisit Lemmas 2–4, including the missing (e(H^*)) factor in Lemma 4, and provide complete proofs

3/How sensitive are the estimated parameters and inferred hidden states to initialization, observation noise, and the number of inspections? Can the authors provide confidence intervals or parameter-recovery experiments on synthetic data?

1 comment

Q1. We thank the reviewer for this observation. In the revision, we fixed this error, and now Equations (9)-(10) include negative emergence terms (Section 3.3). Also, we re-derived the equilibrium in Lemma 4, re-ran all experiments, and updated all related figures. The revised results show comparable performance, confirming that the original conclusions hold.

For the reviewer’s convenience, the previously reported accuracy is also reported in Appendix A.9.

Q2. We added the proofs for Lemmas 2-4 in Appendix A.3-A.5. We also updated Lemma 4 to match the revised equation set.

Q3. We added a parameter-recovery experiment on synthetic data in Appendix A.6. Five of the seven parameters are recovered within about 11% relative error, and the re-fitted model reproduces the synthetic trajectories very precisely (RMSE ≈0.0003\approx 0.0003 frames).

Our response to “Flag For Ethics Review”: Our data come from managed honey bee colonies under standard beekeeping practice. The study involves no human subjects or vertebrate animals, so it does not require IRB review.

Comments on the paper

1 comment

We thank the reviewers for their feedback. All changes are marked in red in the updated PDF. We address each comment below.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.