Appendix B. Benchmarks and Datasets

This appendix curates benchmarks and datasets that enable fair, reproducible evaluation of optimization algorithms across continuous, combinatorial, geospatial, and machine-learning contexts. It explains why standardized test functions and problem instances matter, outlines typical properties used to stress algorithms (multimodality, convexity, separability, differentiability), and points to practical ways to obtain or synthesize data for experimentation. The material is geared toward helping practitioners select representative tasks, generate instances, and measure progress consistently.

For continuous optimization, it highlights canonical test functions used to probe solver behavior. Examples include Ackley (flat plateaus with a central basin), Bohachevsky (smooth, convex, separable), Bukin (rugged ridge with many local minima, non-differentiable, favoring derivative-free search), Gramacy & Lee (1-D with mixed local/global structure), and Griewank (many widespread local minima). By contrasting characteristics—convex versus non-convex, separable versus non-separable, multi- versus unimodal—the appendix motivates when gradient-based methods suffice and when black-box metaheuristics are preferable. It also notes that implementations can be coded from scratch or accessed through common Python libraries such as DEAP, pymoo, and PySwarms.

For combinatorial optimization, the appendix compiles widely used benchmark families—TSP, VRP, job-shop scheduling, assembly-line balancing, quadratic assignment, knapsack, set covering, and bin packing—so algorithms can be compared on standard instances. It then turns to geospatial data, describing spatial attributes and variables relevant to real-world planning (locations, infrastructure, demographics, mobility, environment, and emergency response) and illustrating how open sources and APIs such as OpenStreetMap, Overpass, and elevation services can be tapped for optimization inputs. Finally, for machine-learning-driven and neural combinatorial optimization, it notes that “datasets” are often generated problem instances, with examples including convex-hull point sets, TSP instances for sequence-model training, and large mobility datasets like taxi trip records; accompanying examples show how to prepare these for training and evaluation.

• ALBP: dataset for assembly line balancing problem covered in Chapter 6.
• AdministrativeBoundaries: administrative boundaries of different regions of interest around the world in geojason format that can be used for map visualization.
• BikeShare: Bike Share Toronto (TBS) ridership data that contains anonymized trip data, including: trip start day and time, trip end day and time, trip duration, trip start station, trip end station and user type (see Listing B.4).
• CanadaTraffic: data from Statistics Canada’s dataset for Motor Vehicle Collisions in 2018 such as fatalities per 100,000 population, fatalities per Billion vehicles-kilometres, injuries per Billion vehicles-kilometres, fatalities per 100,000 licensed drivers and injuries per 100,000 licensed drivers (see Listing A.2).
• OntarioHealth: health region of Ontario, Canada (see Listing A.2).
• Police: Toronto Police Service Public Safety data (see Listing A.2).
• PoliticalDistricting: Political districting data used in Chapter 8.
• PtrNets: convex hull and tsp data used by pointer networks (Chapter 11).
• TSP: travelling salesman problem instances.

FAQ