AI Systems Landscape

Optimisation / Operations Research AI — Interactive Architecture Chart

A comprehensive interactive exploration of Optimisation AI — the solve pipeline, 8-layer stack, problem classes, LP/MIP/CP solvers, benchmarks, market data, and more.

~38 min read · Interactive Reference

Hameem M Mahdi, B.S.C.S., M.S.E., Ph.D. · 2026

Senior Principal Applied Scientist | Private Equity Leader | AI Innovative Solutions

📄 Forthcoming Paper

Optimisation Solve Pipeline

The end-to-end workflow for solving optimisation problems — from problem definition through deployment and continuous re-optimisation.

1. Problem Definition

Identify objectives, constraints, decision variables

2. Mathematical Formulation

Model as LP, MIP, CP, or nonlinear program

3. Solve

Execute solver engine; find optimal/near-optimal solution

4. Validate

Check feasibility, sensitivity analysis, stress-test

5. Deploy

Embed in decision systems, APIs, dashboards

6. Monitor / Re-Optimise

Track KPIs, retrigger on data changes

Did You Know?

1

UPS's ORION routing system saves 100 million miles driven and 10 million gallons of fuel per year.

2

Modern LP solvers like Gurobi can solve problems with millions of variables in seconds.

3

The travelling salesman problem — a classic OR challenge — has been exactly solved for up to 85,900 cities.

Knowledge Check

Test your understanding — select the best answer for each question.

Q1. What does LP stand for in operations research?

Q2. What is the travelling salesman problem (TSP)?

Q3. What does "constraint satisfaction" mean?

The Optimisation AI Stack — 8 Layers

Click any layer to expand details about the components and technologies at each level.

8Decision Interface
Dashboards, what-if scenario tools, interactive constraint editors, approval workflows, and human-in-the-loop decision support interfaces.
7Solution Delivery
API endpoints, scheduled batch runs, event-driven triggers, integration with ERP/WMS/TMS, solution serialisation, and result distribution.
6Post-Processing
Solution validation, sensitivity analysis, infeasibility diagnosis, rounding heuristics, warm-start caching, and KPI computation from solver output.
5Solver Engine
LP simplex/interior-point, branch-and-bound/cut (MIP), constraint propagation (CP), heuristic/metaheuristic engines, parallelism, and presolve techniques.
4Model Formulation
Algebraic modelling languages (AMPL, Pyomo, JuMP), objective functions, constraint definition, variable typing, model validation, and reformulation techniques.
3Data & Parameters
Cost matrices, demand forecasts, capacity tables, distance/time matrices, scenario parameters, and data quality pipelines feeding the optimisation model.
2Prediction Layer (ML Forecasting)
Demand forecasting, price prediction, travel-time estimation, failure probability models — ML outputs that serve as parameters to the optimisation model.
1Problem Identification
Business requirement gathering, feasibility assessment, problem classification (LP vs MIP vs CP), complexity estimation, and scope definition.

Optimisation Sub-Types

The major classes of optimisation problems, each with distinct mathematical properties and solution methods.

LP

Linear Programming (LP)

Continuous decision variables with linear objectives and constraints. Solved in polynomial time via simplex or interior-point methods. Foundation of OR.

MIP

Mixed-Integer Programming (MIP)

LP with some integer/binary variables. Branch-and-bound/cut solvers. Core of scheduling, logistics, and facility location.

CP

Constraint Programming (CP)

Domain propagation and backtracking search over finite domains. Excels at scheduling, configuration, and highly combinatorial feasibility problems.

Convex

Convex Optimisation

Minimise convex functions over convex sets. Guarantees global optimum. Includes quadratic, conic, and semidefinite programming.

Combinatorial

Combinatorial Optimisation

Discrete search over finite but exponentially large solution spaces. TSP, graph colouring, bin packing, assignment problems.

Network

Network Optimisation

Shortest path, max flow, min-cost flow, network design. Efficient specialised algorithms (Dijkstra, Ford-Fulkerson).

Stochastic

Stochastic Optimisation

Decision-making under uncertainty with probabilistic parameters. Two-stage, multi-stage, chance-constrained, and robust formulations.

Multi-Obj

Multi-Objective Optimisation

Simultaneous optimisation of conflicting objectives. Pareto front enumeration, weighted-sum, epsilon-constraint methods.

Dynamic

Dynamic Optimisation

Sequential decisions over time with evolving state. Real-time re-optimisation, rolling-horizon, and online optimisation.

Robust

Robust Optimisation

Worst-case optimisation over uncertainty sets. Immunises solutions against parameter perturbations without probability distributions.

Core Architectures & Methods

The algorithmic foundations underlying optimisation solvers and approaches.

Linear Programming (LP)

Methods: Simplex, interior point

Continuous linear objectives and constraints. Polynomial-time solvable. Backbone of supply chain, blending, and resource allocation.

Mixed-Integer Programming (MIP)

Methods: Branch-and-bound, branch-and-cut

Discrete variables for yes/no decisions. Scheduling, logistics, facility location. NP-hard but modern solvers handle millions of variables.

Constraint Programming (CP)

Methods: Domain propagation, backtracking search

Scheduling, configuration, rostering. Excels where feasibility is the core challenge and constraints are highly combinatorial.

Convex Optimisation

Methods: Gradient descent, interior point

Guaranteed global optimum for convex problems. Quadratic, conic, and semidefinite programming. Portfolio optimisation, signal processing.

Nonlinear Programming (NLP)

Methods: Sequential quadratic programming (SQP)

Non-convex objectives/constraints. Engineering design, chemical process optimisation. Local optima challenge.

Heuristic / Metaheuristic

Methods: Local search, simulated annealing, tabu search, genetic algorithms

Large-scale approximate solutions where exact methods are intractable. Vehicle routing, scheduling, layout optimisation.

Dynamic Programming (DP)

Methods: Bellman recursion, memoisation

Sequential decision problems with overlapping subproblems. Inventory management, shortest paths, resource allocation over time.

Leading Solvers & Tools

The major solvers, modelling languages, and platforms powering modern optimisation.

ToolVendorKey Differentiator
GurobiGurobiIndustry-leading LP/MIP solver; fastest commercial solver
IBM CPLEXIBMEnterprise LP/MIP/CP; constraint optimisation at scale
FICO XpressFICOIntegrated solver + modelling for enterprise OR
MOSEKMOSEKBest-in-class conic/semidefinite optimisation
Google OR-ToolsGoogleOpen-source vehicle routing, scheduling, constraint solving
HiGHSOpen-sourceHigh-performance open LP/MIP solver; Linux Foundation
SCIPZuse InstituteLeading academic MIP/MINLP solver
PyomoSandiaPython optimisation modelling language
PuLPPythonSimple LP/MIP Python interface
CVXPYStanfordDisciplined convex programming in Python
JuMPJuliaAlgebraic modelling in Julia; solver-agnostic
AMPLAMPLIndustry algebraic modelling language

Industry Use Cases

How Optimisation / OR AI drives measurable impact across logistics, manufacturing, and supply chain.

Vehicle Routing (VRP)
ExampleDetail
UPS ORIONSaves 100M+ miles/year optimising delivery routes for 100k+ drivers
Amazon RoutingReal-time route optimisation across last-mile delivery network
Google Cloud ORCloud Fleet Routing API for scalable vehicle routing as a service
Warehouse & Facility Location
ExampleDetail
AmazonFulfilment centre placement optimising delivery speed and cost
WalmartDistribution network design covering 4,700+ US stores
DHLGlobal facility placement balancing service levels and cost
Inventory Optimisation
ExampleDetail
Blue YonderAI-driven inventory planning across retail and manufacturing
o9 SolutionsIntegrated demand-supply planning with optimisation engine
RELEX SolutionsRetail inventory and supply chain optimisation platform
Production Scheduling
ExampleDetail
SiemensManufacturing execution with advanced scheduling optimisation
SAP APOAdvanced Planning & Optimisation for production and distribution
AspenTechProcess industry scheduling for chemicals, oil & gas
Supply Chain Network Design
ExampleDetail
LLamasoft (Coupa)End-to-end supply chain network modelling and optimisation
KinaxisConcurrent planning with optimisation across supply chain tiers
Load Planning & Container Packing
ExampleDetail
ORTEC3D load building and vehicle/container packing optimisation
FarEyeLast-mile logistics with load optimisation and route planning

Benchmarks & Performance

Solver performance comparisons and standard benchmark suite coverage.

Solver Performance (Relative)

Standard Benchmark Coverage

Market & Adoption Data

Market sizing and growth projections for the Optimisation / OR AI ecosystem.

Market Segments 2024 ($ Billions)

OR Market Growth 2024 → 2030 (CAGR 16%)

Risks & Limitations

Key challenges and pitfalls in deploying Optimisation / OR AI systems.

Computational Intractability

NP-hard problems may be computationally infeasible at scale; solver time can explode with small increases in problem size.

Model Fidelity

Mathematical models inevitably oversimplify reality; critical real-world constraints may be omitted or linearised away.

Data Quality

Garbage parameters produce garbage solutions — cost coefficients, demand forecasts, and capacity data must be accurate.

Uncertainty

Deterministic models ignore real-world stochasticity; solutions may be fragile under parameter deviations.

Scalability Cliffs

Small growth in problem dimensions can cause exponential growth in solve time, creating sudden performance cliffs.

Objective Misalignment

Optimising the wrong objective function perfectly still yields the wrong answer — objective design is critical and often underinvested.

Key Terminology Glossary

Essential Optimisation / Operations Research AI terminology.

Branch-and-BoundTree-search algorithm that partitions the solution space and prunes subproblems using LP relaxation bounds.
Branch-and-CutExtension of branch-and-bound that adds cutting planes at tree nodes to tighten LP relaxations.
Combinatorial OptimisationOptimisation over a discrete, finite but exponentially large set of feasible solutions.
ConstraintA mathematical condition that a feasible solution must satisfy (equality, inequality, or logical).
Constraint ProgrammingParadigm that finds feasible assignments to variables subject to a set of constraints using propagation and search.
Convex OptimisationMinimisation of a convex function over a convex feasible set; any local optimum is a global optimum.
Decision VariableA variable whose value the solver determines; represents a decision to be made (e.g., quantity, route, schedule).
Dual ProblemA companion problem to the primal whose optimal value provides bounds; strong duality links their solutions.
Dynamic ProgrammingMethod solving complex problems by breaking them into overlapping subproblems via Bellman's principle of optimality.
Feasible RegionThe set of all points satisfying every constraint simultaneously; the solver searches within this space.
HeuristicA practical algorithm that finds good (but not provably optimal) solutions quickly for hard problems.
Integer ProgrammingOptimisation where some or all decision variables must take integer values; core of scheduling and logistics.
Large Neighbourhood SearchMetaheuristic that repeatedly destroys and repairs parts of a solution to explore large neighbourhoods efficiently.
Linear ProgrammingOptimisation of a linear objective subject to linear constraints over continuous variables; solvable in polynomial time.
LP RelaxationRelaxing integer constraints to continuous; provides bounds used by branch-and-bound to prune the search tree.

Visual Infographics

Animation infographics for Optimisation / Operations Research AI — overview and full technology stack.

Regulation

Detailed reference content for regulation.

Regulation & Governance

Optimisation AI intersects with regulation primarily when it makes or informs high-stakes decisions:

Domain Regulatory Considerations
Energy Markets Dispatch optimisation subject to energy market rules and grid reliability standards (FERC, ENTSO-E)
Financial Services Portfolio optimisation and trading algorithms subject to market regulation (SEC, FCA, MiFID II)
Transportation Route and schedule optimisation must comply with driver hour regulations, safety rules, and environmental law
Healthcare Treatment planning optimisation (e.g., radiotherapy) subject to medical device regulation (FDA, MDR)
Labour / Workforce Staff scheduling must comply with labour laws, equity regulations, and union agreements
EU AI Act Optimisation systems used in high-risk domains (employment, critical infrastructure) subject to documentation and transparency requirements
Antitrust Collaborative optimisation between competitors (e.g., shared logistics) may raise competition law concerns

Deep Dives

Detailed reference content for deep dives.

Combinatorial Optimisation — Deep Dive

Classic Problems

Problem Description Complexity Applications
Travelling Salesman Problem (TSP) Find the shortest route visiting all cities exactly once and returning to start NP-hard Delivery routing, PCB drilling, DNA sequencing
Vehicle Routing Problem (VRP) Route a fleet of vehicles to serve customers at minimum cost NP-hard Logistics, last-mile delivery, field service
Bin Packing Pack items into bins using the fewest bins possible NP-hard Container loading, memory allocation, cutting
Knapsack Problem Select items to maximise value within a weight/capacity limit NP-hard Budget allocation, cargo loading, project selection
Job-Shop Scheduling Schedule jobs on machines to minimise makespan or tardiness NP-hard Manufacturing, operating room scheduling
Graph Colouring Assign colours to vertices so no adjacent vertices share a colour NP-hard Register allocation, frequency assignment, scheduling
Facility Location Choose where to open facilities and assign customers to minimise total cost NP-hard Warehouse placement, hospital location, retail
Set Cover / Covering Find minimum-cost set of resources that cover all requirements NP-hard Crew scheduling, sensor placement, test coverage

Modern Approaches to Large-Scale Combinatorial Problems

Approach Description
Branch-and-Cut Combine branch-and-bound with cutting planes to tighten LP relaxations
Column Generation Solve large LP/MIP by generating promising variables (columns) on demand
Benders Decomposition Decompose large problems into a master problem and subproblems
Large Neighbourhood Search (LNS) Iteratively destroy and repair portions of a solution; state-of-the-art for VRP
Neural Combinatorial Optimisation Train neural networks to generate or improve solutions (attention-model TSP solvers)
Hybrid ML + OR Use ML to predict solver parameters, warm-start solutions, or learn branching rules

Constraint Programming — Deep Dive

Core Concepts

Concept Description
Variable A decision to be made; has a domain of possible values
Domain The set of possible values for a variable (integer range, set of options)
Constraint A relation that must hold between variables (e.g., x != y, start + duration <= deadline)
Propagation Systematically reduce variable domains by enforcing constraints
Search Explore reduced search space via intelligent backtracking
Global Constraints High-level constraints with efficient propagation (AllDifferent, Cumulative, Circuit)

When to Use CP vs. MIP

Criterion Constraint Programming Mixed-Integer Programming
Problem Type Scheduling, sequencing, feasibility Resource allocation, network design, planning
Constraint Style Complex logical, disjunctive, global Linear and piecewise linear
Objective Feasibility or optimisation Optimisation with provable bounds
Strengths Rich constraint expressiveness; fast feasibility Tight LP relaxations; strong optimality proofs
Weaknesses Weaker objective bounds Difficulty with complex logical constraints
Best Practice Many modern solvers (CP-SAT, CPLEX) combine both approaches

ML-Augmented Optimisation — Deep Dive

The convergence of machine learning and operations research is one of the most impactful trends in applied AI.

Integration Patterns

Pattern Description Example
Predict-then-Optimise Use ML to forecast inputs, then solve deterministic optimisation with those predictions Forecast demand (ML) then optimise inventory (MIP)
ML for Solver Acceleration Train ML models to learn branching rules, cut selection, or node selection inside solvers Learn2Branch, Ecole framework
ML for Solution Initialisation Use ML to predict warm-start solutions for the solver Predict initial facility locations, then optimise
End-to-End Differentiable Optimisation Embed optimisation layers inside neural networks for joint training OptNet, cvxpylayers, differentiable LP/QP layers
Neural Combinatorial Solvers Train neural networks to directly output solutions to combinatorial problems Pointer Networks, Attention Model for TSP/VRP
Reinforcement Learning for Optimisation Use RL to learn construction or improvement heuristics RL for vehicle routing, scheduling, bin packing
Surrogate-Assisted Optimisation Replace expensive simulations with ML surrogates to speed up optimisation loops Bayesian optimisation, simulation-based design

Key Research

System / Paper Contribution
Ecole (Gasse et al.) OpenAI Gym-like environment for learning to configure and accelerate MIP solvers
OptNet (Amos & Kolter) Differentiable QP solver as a neural network layer
Attention Model (Kool et al.) Transformer-based model that learns to solve routing problems via attention
Google Chip Design (Mirhoseini et al.) RL-based approach to chip floorplanning; published in Nature
DeepMind for MIP (Nair et al.) Neural diving and branching for accelerating MIP solving

Overview

Detailed reference content for overview.

Definition & Core Concept

Optimisation and Operations Research (OR) AI is the branch of artificial intelligence and applied mathematics focused on systems that find optimal or near-optimal solutions to well-defined mathematical problems subject to constraints. It answers the question: given limited resources, conflicting objectives, and operational constraints, what is the best possible decision?

OR is one of the oldest and most impactful computational disciplines, predating modern AI by decades. It was formalised during World War II to optimise military logistics and has since become the backbone of modern supply chains, airline operations, telecommunications networks, financial portfolio management, and manufacturing scheduling. The convergence of OR with machine learning — what is now called ML-augmented optimisation or decision intelligence — represents one of the most impactful frontiers of applied AI.

Unlike predictive AI (which estimates outcomes), generative AI (which creates content), or reinforcement learning (which learns policies through interaction), optimisation AI solves explicitly formulated mathematical programs — finding the values of decision variables that minimise or maximise an objective function subject to constraints.

Dimension Detail
Core Capability Optimises — finds the best feasible solution to a mathematically formulated problem with constraints
How It Works Formulate objective + constraints; apply exact solvers (LP/MIP) or heuristics to find optimal/near-optimal solutions
What It Produces Optimal decisions: schedules, routes, allocations, assignments, plans, and configurations
Key Differentiator Solves explicitly formulated mathematical problems — not learned from data, not discovered through interaction

Optimisation AI vs. Other AI Types

AI Type What It Does Example
Optimisation / Operations Research AI Finds optimal solutions to constrained mathematical problems Vehicle routing, supply chain planning, scheduling, portfolio optimisation
Agentic AI Pursues goals autonomously using tools, memory, and planning Research agent, coding agent, autonomous workflow
Analytical AI Extracts insights and explanations from existing data Dashboard, root-cause analysis, anomaly detection
Autonomous AI (Non-Agentic) Operates independently within fixed boundaries without human input Autopilot, auto-scaling, algorithmic trading
Bayesian / Probabilistic AI Reasons under uncertainty using probability distributions Clinical trial analysis, A/B testing, risk modelling
Cognitive / Neuro-Symbolic AI Combines neural learning with symbolic reasoning LLM + knowledge graph, physics-informed neural net
Conversational AI Manages multi-turn dialogue between humans and machines Customer service chatbot, voice assistant
Evolutionary / Genetic AI Optimises solutions through population-based search inspired by natural selection Neural architecture search, logistics scheduling
Explainable AI (XAI) Makes AI decisions understandable to humans SHAP explanations, LIME, Grad-CAM
Generative AI Creates new original content from learned distributions Write an essay, generate an image, synthesise a video
Multimodal Perception AI Fuses vision, language, audio, and other modalities GPT-4o processing image + text, AV sensor fusion
Physical / Embodied AI Acts in the physical world through sensors and actuators Autonomous vehicle, robot arm, drone
Predictive / Discriminative AI Classifies or forecasts from historical patterns Fraud score, churn probability, demand forecast
Privacy-Preserving AI Trains and runs AI without exposing raw data Federated hospital models, differential privacy
Reactive AI Responds to current input with no learning or memory Chess engine, rule-based spam filter
Recommendation / Retrieval AI Surfaces relevant items from large catalogues based on user signals Netflix suggestions, Google Search, Spotify playlists
Reinforcement Learning AI Learns optimal behaviour from reward signals via trial and error AlphaGo, robotic locomotion, RLHF
Scientific / Simulation AI Solves scientific problems and models physical systems AlphaFold, climate simulation, molecular dynamics
Symbolic / Rule-Based AI Reasons over explicit rules and knowledge to derive conclusions Medical expert system, legal reasoning engine

Key Distinction from Reinforcement Learning: RL learns a policy through trial-and-error interaction with an environment. Optimisation AI solves a known mathematical formulation — the objective and constraints are specified explicitly, not discovered through experience.

Key Distinction from Evolutionary AI: Evolutionary AI is one class of metaheuristic optimisation methods (population-based, biologically inspired). OR encompasses the full spectrum: exact solvers, mathematical programming, heuristics, metaheuristics, and ML-augmented methods.

Key Distinction from Predictive AI: Predictive AI estimates what will happen. Optimisation AI decides what to do about it. In practice, they are often combined: predict demand (Predictive AI), then optimise inventory (Optimisation AI).