Paper: arXiv 2510.02910
Authors: Kevin Kamm
Abstract
This paper studies a joint stochastic optimal control and stopping (JCtrlOS) problem motivated by aquaculture operations, where the objective is to maximize farm profit through an optimal feeding strategy and harvesting time under stochastic price dynamics. We introduce a simplified aquaculture model capturing essential biological and economic features, distinguishing between biologically optimal and economically optimal feeding strategies. The problem is formulated as a Hamilton-Jacobi-Bellman variational inequality and corresponding free boundary problem. We develop two numerical solution approaches: First, a finite difference scheme that serves as a benchmark, and second, a Physics-Informed Neural Network (PINN)-based method, combined with a deep optimal stopping (DeepOS) algorithm to improve stopping time accuracy. Numerical experiments demonstrate that while finite differences perform well in medium-dimensional settings, the PINN approach achieves comparable accuracy and is more scalable to higher dimensions where grid-based methods become infeasible. The results confirm that jointly optimizing feeding and harvesting decisions outperforms strategies that neglect either control or stopping.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematics including Hamilton-Jacobi-Bellman variational inequalities, free boundary problems, and high-dimensional numerical schemes (finite difference and PINNs). It includes a code repository and data, demonstrating empirical implementation, though the model is a simplified ’toy’ aquaculture application rather than a full backtest on real financial data.
Research Flowchart
flowchart TD A["Research Goal: Find optimal feeding & harvesting strategy<br>maximizing aquaculture profit under uncertainty"] --> B["Develop Aquaculture Model<br>- Stochastic price dynamics<br>- Biological growth process<br>- Cost & revenue functions"] B --> C["Formulate JCtrlOS Problem<br>- HJB Variational Inequality<br>- Free Boundary Problem<br>- Control (Feeding) & Stopping (Harvest)"] C --> D["Numerical Solution Methods"] D --> E["Finite Difference Scheme<br>- Grid-based solver<br>- Benchmark for accuracy"] D --> F["Physics-Informed Neural Network<br>- PINN + Deep Optimal Stopping<br>- Scalable to higher dimensions"] E & F --> G["Computational Experiments<br>- Compare JCtrlOS vs. Suboptimal Strategies<br>- Vary dimensions & parameters"] G --> H["Key Findings & Outcomes"] H --> H1["Joint optimization outperforms<br>single or no-control strategies"] H --> H2["Finite differences excel in<br>medium-dimensional settings"] H --> H3["PINN approach achieves<br>comparable accuracy & scalability"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a joint stochastic optimal control and stopping model for aquaculture operations and develops two numerical solution approaches: a finite difference benchmark and a Physics-Informed Neural Network (PINN) combined with a deep optimal stopping algorithm.
Source sentence
This paper studies a joint stochastic optimal control and stopping (JCtrlOS) problem motivated by aquaculture operations, where the objective is to maximize farm profit through an optimal feeding strategy and harvesting time under stochastic price dynamics. We develop two numerical solution approaches: First, a finite difference scheme that serves as a benchmark, and second, a Physics-Informed Neural Network
Assumes 6
- There exists an optimal deterministic biological feeding strategy f(t) that depends only on time. model
Source sentence
We will assume that there exists an optimal (in the sense of maximal growth) deterministic feeding strategy f : R+ →R+ only dependent on time.
- The price dynamics of feed and biomass follow geometric Brownian motions under a risk‑neutral measure. market
Source sentence
We will assume that the price dynamics are both by two geometric Brownian motions with dynamics under a risk-neutral measure Q given by dP F t = rP F t dt + σF P F t dW F t =: bpF (t, P F t )dt + σpF (t, P F t )dW F t , P F 0 = pF > 0, dP B t = rP B t dt + σBP B t dW B t =: bpB(t, P B t )dt + σpB(t, P B t )dW B t , P B 0 = pB > 0, 2We choose a quadratic penalty term for simplicity to penalize over- and underfeeding
- Fish mortality follows an exponential growth model with intrinsic mortality µ and additional mortality µF due to under‑ or overfeeding. model
Source sentence
We assume that the dynamics of hf,u t is given by the following exponential growth model assuming that there is no reproduction of the fish in the farm dhf,u t = −µ −µF (ft −ut)2 hf,u t dt =: bh(t, hf,u t ; ft, ut)dt, hf,u 0 = h0 > 0, where µ > 0 is the intrinsic mortality rate of the fish and µF > 0 the mortality increase due to under- or overfeeding.
- A quadratic penalty term is used to penalize over‑ and underfeeding equally. model
Source sentence
We will assume that the price dynamics are both by two geometric Brownian motions with dynamics under a risk-neutral measure Q given by dP F t = rP F t dt + σF P F t dW F t =: bpF (t, P F t )dt + σpF (t, P F t )dW F t , P F 0 = pF > 0, dP B t = rP B t dt + σBP B t dW B t =: bpB(t, P B t )dt + σpB(t, P B t )dW B t , P B 0 = pB > 0, 2We choose a quadratic penalty term for simplicity to penalize over- and underfeeding
Show 2 more
- The control space is discretized and the maximum over the discrete set of controls is taken. execution
Source sentence
In our numerical implementation, we will discretize the control space as well and take the maximum over the discrete set of controls.
- The neural network training uses a fuzzy boundary sampling with ε = 0.01 to improve stopping region accuracy. model
Source sentence
We chose ε = 0.01 for the fuzzy boundary sampling, which seems to perform best in our experiments. Results.
Data
- Python code and data sets for numerical experiments synthetic public
Source sentence
Keywords: Stochastic Optimal Control, Optimal Stopping, Deep Learning, PINN, Finite Differences, HJB, Numerics, Aquaculture Code availability: The Python code and data sets to produce the numerical experiments are available at https://github.com/kevinkamm/JointOptimalCtrlStopping_Aquaculture.
Costs
Not applicable
Finds 5
- Finite difference methods perform well in medium‑dimensional settings, serving as a benchmark. performance
Source sentence
Numerical experiments demonstrate that while finite differences perform well in medium-dimensional settings, the PINN approach achieves comparable accuracy and is more scalable to higher dimensions where grid-based methods become infeasible.
- The PINN approach achieves comparable accuracy to finite differences and scales to higher dimensions where grid‑based methods become infeasible. performance
Source sentence
Numerical experiments demonstrate that while finite differences perform well in medium-dimensional settings, the PINN approach achieves comparable accuracy and is more scalable to higher dimensions where grid-based methods become infeasible.
- Jointly optimizing feeding and harvesting decisions outperforms strategies that neglect either control or stopping. theoretical
Source sentence
The results confirm that jointly optimizing feeding and harvesting decisions outperforms strategies that neglect either control or stopping.
- The optimal stopping time from the joint problem is later than the optimal stopping time when using the optimal control alone. theoretical
Source sentence
It is noteworthy that the optimal stopping time from the joint optimal control and stopping problem is later than the optimal stopping time when using the optimal control till time τ 1.
- All three deep learning control approaches yield slightly lower mean farm values than the finite difference solution, suggesting possible issues with stopping time determination or control network refinement. negative
Source sentence
However, all three approaches lead to slightly lower farm values than the finite difference approach, which suggests that the either the optimal stopping time is not determined correctly by the PINNs or the control network needs further refinement.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (9)
- On the impact of biological risk in aquaculture valuation and decision making (2025) — link
- On the impact of feeding cost risk in aquaculture valuation and decision making — → 2309.02970
- Approximation Schemes for Mixed Optimal Stopping and Control Problems with Nonlinear Expectations and Jumps (2019) — link
- A Weak Dynamic Programming Principle for Combined Optimal Stopping/Stochastic Control with $\mathcal E ^ f $-expectations (2016) — link
- DGM: A deep learning algorithm for solving partial differential equations (2018) — link
- Numerical Methods for Stochastic Control Problems in Continuous Time (2001) — link
- Neural optimal stopping boundary (2025) — link
- Deep optimal stopping (2020) — link
and 1 more
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations (2019) — link
1 of 9 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.