Machine Learning and Optimization
Machine learning and optimization are complementary computational approaches that address different problems. Machine learning uses data to identify patterns, estimate relationships, and generate predictions, often through optimization algorithms that determine model parameters. Optimization, in turn, identifies decisions that best satisfy specified objectives and constraints. In practical applications, the two approaches can be combined: machine learning estimates uncertain quantities, while optimization uses those estimates to select suitable actions.
The central challenge is not simply to construct a model that fits historical observations. It is to determine whether the model produces useful and sufficiently reliable results when applied to new data and operating conditions.
This resource introduces the principal learning methods, their relationships with statistical modeling, and the practical requirements for developing and evaluating machine learning systems.
Discuss a Machine Learning or Optimization Project
If you are exploring predictive modeling, developing a machine learning application, or seeking to improve operational decisions through mathematical optimization, we can examine the problem, available data, relevant constraints, and possible computational approaches.
An initial discussion can help clarify the objectives, assess technical feasibility, and identify the evidence needed to evaluate alternative solutions.
1. What Is Machine Learning?
Machine learning uses data to estimate model parameters or learn structures that support a specified task. The learning procedure generally minimizes a loss function or optimizes another objective describing the quality of model outputs.
For example, a company may use historical sales, prices, and seasonal information to predict future demand. The algorithm estimates relationships between these variables and sales outcomes.
A model that accurately reproduces its training observations is not necessarily useful. Its performance must also be evaluated on observations not used to construct it.
This distinction between fitting historical data and generalizing to new observations is fundamental to machine learning.
2. Principal Learning Paradigms
Supervised Learning
Supervised learning uses examples containing both input variables and known outcomes. The objective is to learn a relationship that can be applied to new observations. Typical tasks include regression, classification, and forecasting. Common methods include linear and logistic regression, decision trees, random forests, support vector machines, gradient boosting, and neural networks.
For example, a manufacturing company may train a classifier using labeled images of acceptable and defective components. The model then predicts whether newly inspected components belong to either category.
Its suitability depends on more than overall accuracy. False negatives, processing time, changing production conditions, and the consequences of incorrect classifications may be equally important.
Unsupervised Learning
Unsupervised learning investigates structure in data without requiring an explicit outcome label for each observation. Typical applications include clustering, dimensionality reduction, anomaly detection, and representation learning. K-means clustering, principal component analysis, and several neural-network-based methods belong to this broad area.
For example, customer records can be grouped according to purchasing behavior. The resulting groups may support further investigation, but they do not automatically correspond to meaningful market segments.
Their interpretation requires domain knowledge and independent evaluation.
Reinforcement Learning
Reinforcement learning concerns decision-making through interaction with an environment. An agent selects actions and receives rewards or other feedback reflecting their consequences. The objective is typically to develop a policy that maximizes expected cumulative reward over a specified horizon. Applications include robotic control, resource allocation, sequential decision-making, and simulated operational environments.
For example, a warehouse controller might learn inventory replenishment policies through repeated interactions with a simulated supply chain. The quality of the resulting policy depends on the reward specification, training environment, exploration procedure, and correspondence between simulated and operational conditions.
Reinforcement learning is not inherently superior to mathematical optimization, conventional control, or fixed decision rules. The appropriate method depends on the problem.
Evolutionary and Population-Based Methods
Evolutionary algorithms employ mechanisms such as selection, mutation, and recombination to search for candidate solutions. They are primarily optimization methods, although they can be used to learn model parameters, structures, or policies.
Their distinction from supervised or reinforcement learning is important: evolutionary search describes how candidate solutions are improved, not necessarily what information is available to the learner.
3. Machine Learning and Statistical Modeling
Machine learning and statistics share many mathematical foundations, including probability theory, estimation, optimization, and statistical inference.
The distinction between them is not that statistics explains while machine learning predicts. Statistical models can be highly predictive, and machine learning methods can support scientific explanation when used within appropriate research designs.
The more useful distinction concerns the research objective, modeling assumptions, and evidence required.
| Objective | Principal question | Example |
|---|---|---|
| Prediction | How accurately can an outcome be estimated? | Forecast demand |
| Description | What patterns characterize the observations? | Identify customer groups |
| Statistical inference | What can be concluded about population quantities? | Estimate an association |
| Causal inference | What would change under an intervention? | Evaluate a pricing policy |
| Decision optimization | Which feasible action best satisfies an objective? | Allocate inventory |
These objectives may use overlapping techniques but require different validation procedures.
A predictive model identifying a relationship between advertising expenditure and sales does not, by itself, establish the causal effect of increasing advertising expenditure.
Similarly, a model that predicts equipment failures may support maintenance planning without explaining the physical mechanisms responsible for those failures.
Leo Breiman’s influential discussion of statistical modeling emphasized the relationship between data-modeling and algorithmic approaches. Rather than treating them as mutually exclusive, practical research can use both.
4. From Training to Generalization
A conventional supervised learning workflow divides observations into training, validation, and test datasets.
Training data are used to estimate model parameters. Validation data support model selection and configuration. Test data provide an independent assessment after the model-development process is complete.
Cross-validation can improve the efficiency of model comparison, particularly when observations are limited.
However, evaluation procedures must reflect the structure of the data.
For time-series forecasting, random separation of observations may produce misleading results when information from later periods influences earlier predictions. For datasets containing repeated observations of the same individual, machine, or organization, splitting records independently may also overstate generalization performance.
Overfitting and Data Leakage
Overfitting occurs when a model captures patterns specific to the training sample that do not generalize adequately.
Data leakage occurs when information unavailable at the intended prediction time influences model development or evaluation.
For example, standardizing an entire dataset before separating training and test observations can introduce information from the test set into the training process.
Likewise, using revised economic indicators in a historical forecasting experiment may introduce information unavailable to decision-makers at the time.
Both problems can make a model appear more accurate than it is.
5. Measuring Model Performance
Model evaluation should be based on the decision the model is intended to support.
| Application | Possible measures |
|---|---|
| Regression | MAE, RMSE, prediction interval coverage |
| Classification | Precision, recall, F1 score, calibration |
| Forecasting | Forecast error against temporal benchmarks |
| Anomaly detection | False alarm rate, detection rate |
| Industrial vision | False negatives, throughput, inference latency |
| Decision policies | Cumulative reward, constraint violations, operational cost |
No single measure adequately describes performance in every setting.
Consider a defect-detection model achieving 98% classification accuracy. If defective products are rare, the model could attain this result while failing to detect a substantial proportion of actual defects.
A useful evaluation must therefore consider class prevalence, false-negative consequences, and operational requirements.
Performance should also be evaluated under changes in data quality, environmental conditions, and input distributions.
6. Interpretability, Robustness, and Reliability
A model can be accurate but difficult to interpret, or interpretable but insufficiently accurate for a particular task. These properties should be evaluated separately.
Interpretability concerns the extent to which model behavior can be understood. Robustness concerns sensitivity to changes or disturbances. Reliability concerns satisfactory behavior under specified operating conditions. For example, an industrial inspection system may perform well under laboratory lighting but exhibit increased false-negative rates when camera orientation or illumination changes.
The relevant evaluation is not whether the model works in one demonstration. It is whether performance remains acceptable across the defined operating conditions. This distinction is particularly important when machine learning is integrated into workflows that include software tools, operational decisions, or physical equipment.
7. Applications in Business and Industry
Demand Forecasting
Machine learning can estimate future demand using historical orders, prices, seasonality, promotions, and other relevant variables. The predictive model does not determine an inventory policy by itself. Forecasts must be incorporated into decisions about stock levels, replenishment, costs, and service requirements.
Predictive Maintenance
Machine learning can analyze sensor measurements and maintenance records to identify patterns associated with equipment deterioration. Its usefulness depends on sensor quality, the definition of failure events, maintenance constraints, and the operational value of predictions.
Industrial Computer Vision
Image classification and object detection can support quality inspection, identification, and monitoring. Evaluation should include changing imaging conditions, detection errors, processing latency, and integration with existing inspection processes.
Financial Analysis
Machine learning can investigate relationships among economic indicators, market variables, accounting information, and financial outcomes. Applications include forecasting, risk assessment, classification, and systematic strategy research. Financial time series require particular attention to changing market regimes, information availability, transaction costs, and out-of-sample validation.
Organizational and Institutional Modeling
Machine learning can help estimate behavioral relationships, classify simulated agents, or construct approximations of computationally expensive models. In agent-based modeling, such techniques may support parameter estimation, calibration, behavioral approximation, and the analysis of simulation outputs. However, predictive agreement does not establish that the underlying behavioral or institutional mechanisms have been correctly identified.
Discuss a Machine Learning or Optimization Project
If you are exploring predictive modeling, developing a machine learning application, or seeking to improve operational decisions through mathematical optimization, we can examine the problem, available data, relevant constraints, and possible computational approaches.
An initial discussion can help clarify the objectives, assess technical feasibility, and identify the evidence needed to evaluate alternative solutions.
The E-Flow warehouse management system illustrates production experience with tracked goods and operational constraints. Related work on feasibility studies and computational assessment is described under AI Consulting and Computational Systems. The original WMS deployment and subsequent simulation research remain distinct.
Discuss a Machine Learning or Optimization Project
If you are exploring predictive modeling, developing a machine learning application, or seeking to improve operational decisions through mathematical optimization, we can examine the problem, available data, relevant constraints, and possible computational approaches.
An initial discussion can help clarify the objectives, assess technical feasibility, and identify the evidence needed to evaluate alternative solutions.
8. Machine Learning and Computational Optimization
Machine learning and optimization are closely connected, but they solve different problems.
A learning algorithm typically uses optimization to estimate model parameters. A decision system may then use the model’s predictions as inputs to a separate optimization problem.
Consider a distribution network. A machine learning model predicts demand at different locations. An optimization model then determines how much inventory to allocate, subject to transportation capacity, storage constraints, and service requirements. The distinction is:
Learning estimates uncertain quantities. Optimization selects decisions under an objective and constraints.
The two methods can be integrated, but their errors and assumptions should be evaluated separately. This distinction also applies to financial portfolios. Machine learning may estimate expected returns or risk measures. Portfolio optimization then selects asset weights using these estimates and specified constraints. Neither machine learning nor optimization guarantees superior investment performance. The result depends on the reliability of estimates, feasibility of decisions, and conditions encountered after implementation.
9. Simulation and Reinforcement Learning
Simulation provides a controlled environment for examining model behavior when direct operational experiments are expensive or impractical.
For example, a simulated supply chain can represent demand uncertainty, transportation delays, replenishment policies, and disruptions. A reinforcement learning agent may interact with this environment to develop a replenishment policy.
Its performance can then be compared with fixed rules, conventional optimization, or model-predictive control. The comparison should use common scenarios and evaluation criteria, including service levels, costs, constraint violations, and robustness.
A successful result in simulation is evidence about behavior under the simulated assumptions. Operational validity requires further examination of the differences between the model and the real system.
10. A Reproducible Machine Learning Workflow
A practical machine learning project can be organized into six stages.
- Define the task. Specify the required prediction or decision, operating conditions, and acceptance criteria.
- Examine the data. Assess completeness, representativeness, measurement quality, and information availability.
- Establish benchmarks. Compare against simple statistical models, existing procedures, and non-ML alternatives.
- Train and validate. Use appropriate dataset separation, model selection, and performance measures.
- Test robustness. Examine distribution changes, difficult cases, uncertainty, and operational failure conditions.
- Document outcomes. Report the experimental procedure, results, limitations, and requirements for implementation.
The central question is not whether a sophisticated model can be trained, but whether its behavior provides sufficient evidence for the intended use.
11. Classical Optimization
Optimization concerns the selection of decisions that minimize costs, maximize benefits, or satisfy other objectives subject to specified constraints. It provides a mathematical foundation for applications in economics, operations research, engineering, finance, logistics, and artificial intelligence.
The practical importance of optimization lies in making tradeoffs explicit. Resources are limited, objectives may conflict, and decisions must often satisfy technical, financial, or institutional requirements.
This resource introduces the principal optimization concepts, their mathematical foundations, and their applications to computational decision-making and simulation.
12. What Is an Optimization Problem?
An optimization problem consists of three principal elements: decision variables, an objective function, and constraints.
Decision variables describe the choices available to a decision-maker. The objective function evaluates those choices. Constraints define which combinations are admissible. For a minimization problem, the general formulation is:
Minimize f(x), subject to x ∈ D
Here:
- x represents the decision variables.
- f(x) is the objective function.
- D is the feasible set containing decisions that satisfy the constraints.
An optimal solution is a feasible decision whose objective value is no greater than that of any other feasible decision. The same structure applies to maximization problems by reversing the objective.
A Simple Business Example
Suppose a company operates two warehouses and must satisfy demand from three regional markets. The company must decide how many units to ship from each warehouse to each destination. Transportation cost is the objective to minimize. Warehouse capacity and customer demand define the principal constraints. The problem can be written as:
Minimize total transportation cost
Subject to:
- Shipments cannot exceed available warehouse capacity.
- Customer demand must be satisfied.
- Shipment quantities cannot be negative.
Once the costs and capacities are specified, an optimization model can identify a cost-minimizing distribution plan, provided a feasible optimum exists.
13. Feasibility, Optimality, and Mathematical Guarantees
An optimization problem does not necessarily have an optimal solution. The constraints may be mutually inconsistent, producing an empty feasible set. Alternatively, the objective may improve without bound. A finite optimal value may also exist without being attained by any feasible decision.
The existence of a solution therefore depends on the mathematical properties of the objective function and feasible set. For example, a continuous objective function defined on a nonempty, compact feasible set attains both a minimum and a maximum. This is a consequence of the extreme value theorem.
Local and Global Optima
A local optimum is better than or equal to sufficiently nearby feasible alternatives.
A global optimum is better than or equal to every feasible alternative.
The distinction matters because algorithms for nonconvex problems may converge to local solutions without establishing global optimality. For differentiable unconstrained problems, a stationary point may be identified by a zero gradient. However, stationarity alone does not establish that the point is a minimum. Additional conditions or global analysis may be required.
14. Principal Classes of Optimization Problems
Different mathematical structures require different solution methods.
| Problem class | Characteristics | Example |
|---|---|---|
| Linear programming | Linear objective and constraints | Transportation allocation |
| Mixed-integer programming | Some decisions must be integer-valued | Facility location |
| Convex optimization | Convex feasible set and convex minimization objective | Certain portfolio problems |
| Nonlinear optimization | Nonlinear objectives or constraints | Engineering design |
| Combinatorial optimization | Selection among discrete configurations | Vehicle routing |
| Stochastic optimization | Explicit representation of uncertain quantities | Inventory planning |
| Robust optimization | Decisions protected against defined uncertainty sets | Capacity allocation |
| Dynamic optimization | Decisions affect subsequent states | Multi-period production |
The classification is not merely descriptive. It helps determine which algorithms, guarantees, and computational resources are appropriate.
15. Linear Programming
Linear programming minimizes or maximizes a linear objective subject to linear constraints. It is widely used in production planning, transportation, workforce allocation, and resource distribution. Consider a company producing two products. Let x₁ and x₂ denote the quantities manufactured. Suppose their unit contributions are $30 and $20. The objective is:
Maximize profit = 30x₁ + 20x₂
Assume the following resource constraints:
2x₁ + x₂ ≤ 100
x₁ + 2x₂ ≤ 80
x₁ ≥ 0 and x₂ ≥ 0
The first two inequalities represent limits on available resources. The optimal solution is:
x₁ = 40, x₂ = 20
The corresponding objective value is:
Maximum profit = $1,600
Both resource constraints are binding at this solution. The example illustrates how an optimization model converts resource limitations into an explicit allocation decision.
16. Integer and Combinatorial Optimization
Some decisions cannot be represented adequately by continuous variables. A factory may either be opened or remain closed. A vehicle may either be assigned to a route or not assigned. A machine may need to produce an integer number of components. Such problems require discrete decision variables. Mixed-integer programming combines continuous and integer variables within a mathematical formulation. Typical applications include facility location, scheduling, production planning, and supply chain network design.
An Illustrative Facility-Location Problem
Suppose a company is considering three possible distribution centers.
Opening each center creates a fixed operating cost, while serving customers produces transportation expenses. The decision is to identify which centers should operate and how demand should be allocated.
A mathematical model can minimize total cost subject to capacity and service constraints. Because opening a center is a yes-or-no decision, binary variables are appropriate. Many such problems can be solved using mixed-integer programming or constraint programming methods, depending on their mathematical structure.
17. Convexity and Optimality Conditions
Convexity is an important property because it can simplify both analysis and computation. A feasible set is convex when the line segment connecting any two feasible points remains inside the set. For minimization, a convex objective defined over a convex feasible set has the useful property that every local minimum is also a global minimum. When the relevant regularity conditions hold, first-order optimality conditions can help characterize solutions.
Lagrange Multipliers
Lagrange multipliers provide a method for analyzing differentiable optimization problems with equality constraints. The multipliers can also have an economic interpretation as marginal values associated with changes in constraints, under appropriate conditions. For example, in a production problem, a multiplier associated with a capacity constraint may describe the marginal change in optimal profit resulting from a small increase in available capacity.
Karush–Kuhn–Tucker Conditions
The Karush–Kuhn–Tucker (KKT) conditions extend the analysis to problems involving inequality constraints. They provide necessary conditions for local optimality under suitable constraint qualifications. For convex optimization problems, these conditions can also characterize global optimality under appropriate assumptions. They should not be interpreted as universal guarantees that an arbitrary numerical algorithm has found the best possible solution.
18. Duality and Shadow Prices
Many optimization problems admit an associated dual problem. Duality provides mathematical relationships between the original problem and another formulation involving variables often associated with constraints. For a minimization problem, feasible solutions to an appropriate dual formulation provide lower bounds on the primal objective value. Under strong duality, the primal and dual optimal values coincide. Strong duality holds for linear programming when finite optimal solutions exist and under suitable conditions for certain other problem classes.
Economic Interpretation
Consider a company operating at its maximum available production capacity. If additional capacity would increase optimal profit, the corresponding shadow price may quantify the marginal economic value of relaxing that constraint. This can support investment decisions. However, shadow prices are local sensitivity measures under the assumptions of the model. They need not remain constant when resource availability changes substantially.
19. Optimization Under Uncertainty
Many practical decisions must be made before uncertain quantities become known. Demand, prices, transportation times, equipment availability, and financial returns may all vary. Different optimization frameworks represent uncertainty in different ways.
Stochastic Optimization
Stochastic optimization uses probability distributions or scenarios to represent uncertainty. A company may choose inventory levels before demand becomes known and evaluate expected costs across multiple demand scenarios. The result depends on the assumed probability distribution and the structure of the decision problem.
Robust Optimization
Robust optimization evaluates decisions against a defined set of possible parameter values. Instead of relying on a single expected outcome, it seeks solutions that remain feasible or achieve specified performance guarantees across the uncertainty set. Robust solutions may sacrifice some performance under favorable conditions to improve outcomes under adverse conditions.
Multiobjective Optimization
Business and engineering decisions often involve more than one objective. A logistics company may seek to minimize costs, reduce delivery times, and limit environmental impacts. A solution that improves one objective may worsen another. Pareto efficiency provides a useful concept: a feasible solution is Pareto efficient if no other feasible solution improves at least one objective without worsening another. Multiobjective optimization helps identify and compare such tradeoffs rather than assuming that one universal solution is best.
20. Dynamic Optimization and Sequential Decisions
Dynamic optimization concerns decisions made over time, particularly when present actions influence future opportunities. Examples include inventory replenishment, energy storage, production scheduling, investment allocation, and equipment maintenance. A dynamic problem typically includes states, actions, transition rules, and a cumulative objective.
The Bellman principle of optimality provides a basis for decomposing certain sequential decision problems into smaller subproblems. Dynamic programming uses this structure to compute optimal decisions when the necessary assumptions and computational requirements are satisfied. However, the number of possible states and decisions can grow rapidly.
This limitation, commonly called the curse of dimensionality, motivates approximate dynamic programming, reinforcement learning, decomposition methods, and other computational approaches.
21. Optimization, Machine Learning, and Simulation
Optimization, machine learning, and simulation are complementary methods.
Machine learning estimates patterns or relationships from data.
Simulation evaluates how a system behaves under specified mechanisms, conditions, and decisions.
Optimization searches for decisions that satisfy a stated objective and constraints.
Consider an inventory distribution problem.
A machine learning model predicts demand. A simulation model represents orders, delays, stockouts, and replenishment behavior. An optimization model determines shipment quantities or replenishment schedules. The complete system may use all three components, but each contributes different evidence. A reliable implementation requires assessing errors in the predictive model, assumptions in the simulation, and feasibility of the optimization outputs.
Simulation-Based Optimization
Some practical systems are too complex to represent through an analytically convenient objective function. In these cases, candidate decisions can be evaluated through simulation. An optimization procedure then searches for configurations that perform well across repeated simulation experiments.
For example, an agent-based supply chain model may represent customers, suppliers, warehouses, and heterogeneous operating rules.
Simulation-based optimization can compare alternative inventory policies while accounting for interactions that are difficult to express in a conventional mathematical program.
The quality of the results depends on the simulation model’s validity, the experimental design, and the treatment of stochastic variability.
22. Applications in Finance, Industry, and Robotics
Financial Portfolio Allocation
Portfolio optimization allocates capital among assets subject to risk, return, and investment constraints. A classical mean-variance model balances expected return and portfolio variance. However, estimated returns and covariances are uncertain. A mathematically optimal portfolio may be highly sensitive to estimation errors. Practical models therefore often incorporate transaction costs, position limits, turnover constraints, and robustness analysis.
Supply Chain and Distribution
Optimization supports warehouse allocation, inventory replenishment, transportation routing, and production scheduling. Its practical value depends on representing operational constraints accurately. An apparently efficient solution may be infeasible if labor availability, vehicle capacities, delivery windows, or uncertain lead times have been omitted.
Robotics and Control
Optimization is used in trajectory planning, control allocation, task scheduling, and resource management. Model-predictive control repeatedly solves a finite-horizon optimization problem using updated information about the system state. The method can incorporate physical constraints, although its effectiveness depends on the dynamics, computational budget, and model assumptions.
Energy and Infrastructure
Optimization can support energy dispatch, capacity planning, infrastructure maintenance, and service allocation. These applications frequently involve discrete decisions, uncertain demand, and competing economic or operational objectives.
Computational Research and Quantum Optimization
Certain combinatorial problems can be represented as quadratic unconstrained binary optimization (QUBO) problems.
Such formulations are relevant to classical optimization and to experimental quantum algorithms, including the Quantum Approximate Optimization Algorithm (QAOA).
Quantum approaches remain research-dependent alternatives. Representing a problem as a QUBO does not establish that a quantum algorithm will outperform classical methods.
Comparative experiments should evaluate solution quality, computational resources, scaling behavior, and suitable classical baselines.
Discuss a Machine Learning or Optimization Project
If you are exploring predictive modeling, developing a machine learning application, or seeking to improve operational decisions through mathematical optimization, we can examine the problem, available data, relevant constraints, and possible computational approaches.
An initial discussion can help clarify the objectives, assess technical feasibility, and identify the evidence needed to evaluate alternative solutions.
23. Evaluating an Optimization Method
The quality of an optimization method should be assessed through more than its reported objective value. A reproducible evaluation should address:
- Problem specification: Are the objective, decision variables, and constraints clearly defined?
- Feasibility: Does the proposed solution satisfy all required constraints?
- Optimality: Is the solution globally optimal, locally optimal, or an approximation?
- Computational effort: What time, memory, and hardware resources are required?
- Sensitivity: How do results change when parameters or assumptions vary?
- Operational validity: Can the solution be implemented under real conditions?
For mixed-integer optimization, solvers may report an optimality gap comparing the best feasible objective value with a mathematical bound.
This information helps distinguish a proven optimum from a high-quality solution whose optimality has not been established.
For stochastic or simulation-based optimization, repeated experiments and uncertainty estimates may also be necessary.
A good mathematical solution is not automatically a good operational decision if the model omits relevant constraints or uses unreliable data.
24. A Practical Optimization Workflow
An applied optimization project can be organized into six stages.
- Define the decision. Identify the available alternatives and the objectives of the organization.
- Represent constraints. Specify resource limits, operating requirements, and feasibility conditions.
- Select the formulation. Determine whether linear, integer, convex, stochastic, dynamic, or simulation-based methods are appropriate.
- Establish benchmarks. Compare the proposed approach with existing rules and simpler alternatives.
- Solve and validate. Examine feasibility, objective quality, computational performance, and robustness.
- Evaluate implementation. Translate results into operating procedures and identify assumptions requiring monitoring.
The objective is not merely to obtain a numerical solution, but to establish whether that solution is useful for the underlying decision.
References and Further Reading
Optimization:
Sundaram, R. K. (1996). A First Course in Optimization Theory. Cambridge University Press.
Boyd, S., and Vandenberghe, L. (2004). Convex Optimization. Cambridge University Press. Official academic resource.
Bertsimas, D., and Tsitsiklis, J. N. (1997). Introduction to Linear Optimization. Athena Scientific.
Bertsekas, D. P. (2017). Dynamic Programming and Optimal Control, Volume I, 4th edition. Athena Scientific.
Google. OR-Tools: Optimization and Operations Research. Official documentation.
Machine Learning
Breiman, L. (2001). Statistical Modeling: The Two Cultures. Statistical Science, 16(3), 199–231. DOI.
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edition. Springer. DOI.
Sutton, R. S., and Barto, A. G. (2018). Reinforcement Learning: An Introduction, 2nd edition. MIT Press. Publisher.
Scikit-learn Developers. Common Pitfalls and Recommended Practices. Official documentation.
Technical and educational material. Examples describe analytical approaches and do not imply completed deployments or validated performance in particular operating environments.
