Optimization Techniques – Comprehensive Exam Notes
Course Logistics
- Faculty & Department: Accra Technical University – Applied Mathematics & Statistics
- Programme / Course Code: M.Tech Data Science & Industrial Analytics – \text{MDS 616}
- Title / Credits / Schedule: “Optimization Techniques”, 3-credit Saturday course (7 a.m.–10 a.m., KF9)
- Facilitator: Dr. M. O. Amoamah
- Delivery & Grading
- Lectures (core theory + worked examples)
- Assignments / exercises 10 %
- Presentations 10 %
- Mid-semester exam 20 %
- End-semester exam 60 %
Learning Outcomes
- Apply core principles of optimization across engineering, economics & data-science tasks.
- Formulate & solve linear and non-linear models (constrained & unconstrained).
- Deploy specialized sub-fields: integer programming, meta-heuristics, network & assignment models.
- Operate industry software (Excel-Solver, Python/Matlab, etc.) to obtain, analyse & interpret optimal solutions.
Reading Benchmarks
- Bazaraa et al., Linear Programming & Network Flows (5 ed.)
- Nocedal & Wright, Numerical Optimization (2 ed.)
- Boyd & Vandenberghe, Convex Optimization
- Hillier & Lieberman, Introduction to Operations Research (10 ed.)
1. Foundations: What is Optimization?
- Definition: Mathematical process for choosing the best (maximum or minimum) value of an objective function under given constraints.
- Basic Terminology
- Objective Function f(x) → maximise or minimise.
- Decision Vector x=(x1,\dots,xn).
- Constraints gi(x)\le,=,\ge bi.
- Feasible Region = all x satisfying constraints.
- Examples / Industries
- Business profit maximisation, portfolio risk/return, airline crew scheduling, bridge design (weight vs strength), routing & transportation, inventory, project scheduling.
2. Linear Optimization (Part I)
- Standard form (primal):
\begin{aligned}
\text{Optimise}\; &Z = \sum{j=1}^{n} cj xj\
\text{s.t.}\; &\sum{j=1}^{n} a{ij}xj \le (=,\ge)\; bi,\; i=1,\dots,m\
&xj\ge 0\; (\text{or integer/binary})\end{aligned} - Resources b_i interpreted broadly (time, budget, capacity…).
- Steps to formulate
- Identify decision variables.
- Collect numerical data (constants/parameters).
- Translate real-world limits into linear constraints.
- Write linear objective (max profit / min cost).
2.2 Graphical Method (2-variable)
- Plot each inequality as a half-plane; feasible region = polygon intersection.
- Optimal corner = any vertex with best Z (convexity of LP ensures optima on extreme points).
- Procedure bullets: define x,y → list constraints → convert to equalities → draw → shade feasible zone → enumerate corner points → evaluate Z.
2.3 Simplex Algorithm
- Iterative tableau procedure that moves from one basic feasible solution to an adjacent one with improved Z until no positive (max) or negative (min) reduced cost remains.
- Slack / surplus variables convert “≤” or “≥” to equalities.
- Key tableau entries
- C_B: coefficients of current basis.
- Zj, Cj-Z_j: optimality test.
- Ratio column \theta=\dfrac{\text{RHS}}{\text{pivot column}^{+}} selects leaving variable.
- Computational support: Excel Solver → choose Simplex LP.
2.4 Duality & Sensitivity
- Every LP has a dual; dual variables have economic meaning (shadow prices).
- If primal is max, dual is min (and vice-versa).
- Correspondence:
- Primal variables \leftrightarrow Dual constraints.
- Primal RHS b \leftrightarrow Dual objective coefficients.
- Shadow price: change in optimal Z per unit increase in a RHS value (valid within allowable range).
- Post-optimality / sensitivity: analyse how changes in cj, bi, or new variables/constraints affect optimality without re-solving from scratch (via final tableau or Solver sensitivity report).
3. Non-Linear Optimization (Part II)
3.1 Problem Classes
| Type | Objective/Constraints | Example |
|---|
| Linear | Both linear | Resource allocation LP |
| Non-linear | At least one non-linear term | Portfolio variance 0.06x1^2+0.04x2^2+0.04x1x2 |
| Constrained | Equality/inequality present | Beam design h^2w\ge 200 |
| Unconstrained | No constraints on x | Least-squares regression |
3.2 Unconstrained Single-Variable
- Necessary condition: f'(x^*)=0.
- Second-derivative test:
- f''(x^*)>0 → local minimum.
- f''(x^*)<0 → local maximum.
- Graph notions: local/global extrema, inflection points.
3.3 Unconstrained Multivariable
- Gradient \nabla f(x)=\left[\dfrac{\partial f}{\partial x1},\dots,\dfrac{\partial f}{\partial xn}\right].
- Hessian H(x) is matrix of second partials.
- Positive definite → local minimum.
- Negative definite → local maximum.
- Indefinite → saddle point.
- Definiteness tests: leading principal minors (Sylvester criteria).
3.4 Constrained NL Optimization
Lagrange Multipliers (Equality)
- Build L(x,\lambda)=f(x)-\sum{i=1}^{m}\lambdai g_i(x).
- Solve \nablax L=0 & gi(x)=0 simultaneously.
- \lambda_i measures sensitivity of optimum to RHS of constraint.
Kuhn–Tucker (Inequality)
- Introduce slack si\ge0, set up Lagrangian with multipliers \lambdai\ge0.
- Complementary slackness: \lambdai gi(x)=0 at optimum.
3.5 Convexity & Global Optimality
- Set S convex ⇔ any two points x,y\in S imply \theta x+(1-\theta)y\in S for \theta\in[0,1].
- Function convex ⇔ H(x) positive (semi)-definite ∀ x.
- In convex problems any local optimum is global.
3.6 Software (Excel Solver)
- Use GRG Nonlinear (gradient-based) for smooth models.
- Use Evolutionary when model is non-smooth, discontinuous or multi-modal.
- Options: multistart, tolerance, convergence diagnostics.
4. Integer Programming (Part III)
- Variables restricted to integers (0/1, counts, on/off).
- Types
- Pure IP – all variables integer.
- Mixed IP – some integer, others continuous.
- Zero-One (Binary) – variables \in{0,1}.
- Solution Methods
- Branch-and-Bound (systematically explore tree, prune via bounds).
- Cutting Planes (add linear cuts to remove fractional solutions).
- Typical Applications: capital budgeting, facility location, crew assignment, project selection, knapsack, scheduling.
- Why? Exact methods may be infeasible on huge or highly non-convex spaces.
- Core concepts: randomisation, neighbourhood search, acceptance of worse moves to escape local minima.
- Simulated Annealing (SA)
- Temperature parameter T controls acceptance probability \exp\left(-\Delta f/T\right).
- Gradual cooling → convergence to near-global optimum.
- Genetic Algorithm (GA)
- Population encoding, fitness evaluation, selection, crossover, mutation.
- Good for parameter tuning, feature subset selection.
6.1 Transportation Model
- Balanced if \sum ai = \sum bj; else add dummy source/destination with zero cost.
- Formulation variables x{ij} units shipped from i to j; minimise \sum c{ij}x_{ij} subject to row (supply) & column (demand) equalities.
- Initial solution heuristics: North-West Corner, Least-Cost (Minimum-Cost), Vogel Approximation; test optimality via MODI / stepping-stone.
6.2 Assignment Problem
- Special transportation with unit supply/demand; binary variables x_{ij}\in{0,1}.
- Hungarian Method or LP with binary constraints; cost/profit matrix.
6.3 Travelling Salesman Problem (TSP)
- Visit each city once, return to start, minimise total distance.
- LP with subtour-elimination (Miller–Tucker–Zemlin variables u_i) or heuristics (nearest-neighbour, 2-opt, etc.).
6.4 Knapsack (0-1)
- \text{Maximise}\;\sum vi xi
\text{s.t. }\sum wi xi \le W,\; x_i\in{0,1}. - Exact: dynamic programming, branch & bound; heuristic: greedy by vi/wi ratio.
7. Excel-Solver Operational Guide
- Activate add-in: File → Options → Add-ins → Manage Excel Add-ins → Solver.
- Sheet layout: decision variables in cells, formulas for objective & constraints in separate cells.
- Dialog setup: Set Objective (max/min), By Changing variables, Add constraints, choose method (Simplex LP / GRG Nonlinear / Evolutionary).
- Reports:
- Answer (values, binding status).
- Sensitivity (shadow prices, reduced costs, allowable ranges).
- Limits (variable bounds with others fixed).
- Interpretations:
- Binding ⇒ zero slack.
- Reduced cost 0 on basic vars; non-basic vars show improvement needed to enter.
- Shadow price valid within allowable increase/decrease of RHS.
- LP Formulations: data-cleaning vs feature-engineering, ensemble model mix, social-media campaigns, model-training scheduling, forecasting accuracy vs cost, churn-reduction, tech-product mix, text-analytics task assignment, software subscription, experimentation budgeting.
- Non-linear Excel Tasks: portfolio variance minimisation, Cobb-Douglas production, logistic curve fitting, quadratic advertising returns, circular area vs perimeter.
- Unconstrained Multivariate: 10 analytical gradient/Hessian exercises.
- Convexity Proofs: prove convexity & verify optimality at (0,0) for 10 given functions.
- Integer & Network Applications:
- Detailed transportation, assignment, TSP & knapsack datasets (warehouse–store, worker–task, city tours, backpack capacity).
- Students expected to state decision variables → objective → constraints → solve graphically/analytically → confirm with Solver.
- Solver Report Interpretation: understand answer & sensitivity snapshots (final Z, shadow prices, allowable ranges, reduced costs).
9. Ethical & Practical Insights
- Optimal solutions must respect real-world rules (labour laws, safety, budgets).
- Shadow-price info aids negotiation & resource valuation.
- Integer and heuristic methods accommodate indivisibilities & computational hardness prevalent in industry datasets.
10. Quick-Reference Equations
- LP Duality Pair (max primal, min dual)
\begin{aligned}
\max\; &c^T x & \quad \min\; b^T y \
\text{s.t.}&\; Ax \le b & \text{s.t. } A^T y \ge c \
& x\ge 0 & y \ge 0\end{aligned} - Portfolio Variance (two assets)
\sigma^2 = \sigmaA^2 x1^2 + \sigmaB^2 x2^2 + 2\,\text{Cov}{AB}\,x1x2 with x1+x_2=1. - Lagrange First-Order Conditions
\nabla f(x^) - \sum{i}\lambdai \nabla gi(x^) = 0,\quad gi(x^*)=0. - Kuhn-Tucker Complementarity
\lambdai\ge0,\; gi(x^)\le0,\; \lambdai gi(x^)=0.
11. Exam Tips
- Always begin with clear definition of variables; dimensionally check units.
- For LP proofs, sketch feasible space even when not required—visual intuition helps.
- Memorise standard forms & dual conversion rules (≤ ↔ ≥, max ↔ min).
- When using Solver, keep raw coefficients in dedicated cells to leverage sensitivity report.
- For non-linear calculus questions: compute gradient, set =0, then Hessian test.
- In integer models, justify integrality (people, servers, projects) before applying IP tools.
- Interpret every numerical answer (profit, cost, risk) in context—marks awarded for insight, not just numbers.