Synthetic, reproducible example

A worked network-access example

Compare separate routes with a network that shares paths between destinations.

New synthetic teaching companion, not the dissertation implementation or its research benchmark. Costs are invented units, not travel times or distances.

Run it locally

Python 3.10 or newer. Standard library only; no accounts, credentials, or external datasets. From the repository root:

python examples/network-access/demo.py
python -m unittest discover -s examples/network-access -p test_demo.py -v

Read the guide Inspect Python code Inspect tests Raw evaluation report (JSON)

Recorded demonstration results

These are outputs of the included teaching example, not measurements of a client system or field accuracy.

Independent shortest-path union cost: 12
Minimum shared-network cost: 10
Exhaustively checked edge subsets: 32

Why shared edges change a network-access decision — An invented network for learning, not a dissertation test case.

Worked example: independent routes
Target	Route	Route cost
T1	S → T1	6
T2	S → T2	6

Worked example: selected shared-network edges
Edge	Cost counted once
H ↔ S	4
H ↔ T1	3
H ↔ T2	3

Measured synthetic micro-benchmark (environment-specific)
Targets	Edges	Subsets	Path-union cost	Shared cost	Dijkstra median ms	Exhaustive median ms
2	5	32	12	10	0.011	0.0548
3	9	512	18	13	0.0168	1.0468
4	13	8192	24	16	0.0228	21.8929

Interpretation and limitations

For each target the direct path costs 6, versus 7 through H. Yet sharing S–H gives a total selected-edge cost of 4 + 3 + 3 = 10 instead of 6 + 6 = 12. Individual journeys need not become shorter.
The comparable baseline is the union of independent shortest-path edges, counted once. Summing journey costs would double-count any shared edge.
Solver settings: simple undirected graphs; positive integer edge costs; all targets required; optional intermediate nodes; exhaustive enumeration of every subset; no pruning, tolerance, heuristic, random seed, external solver, or time limit; single process; hard cap 16 edges (65,536 subsets).
Timings are medians of 5 runs after one warm-up, measured with perf_counter_ns, including validation and solving but excluding fixture creation, report assembly, and file I/O. They are illustrative, not a scalable algorithm comparison.
Measured environment: CPython 3.12.14, Windows 11, AMD64. Workload and timing noise vary by machine; rerunning overwrites timings.
This familiar minimum connecting-subgraph example is not a novelty claim, proof about the original formulation, or validation of convex-hull preprocessing. The public R source uses a separate OMPR/GLPK formulation with flow variables.
No unpublished manuscript, employer code, client records, or real location data is included. No journal publication or DOI is asserted.

← Back to case study