Plan like a cat: avoid what matters, remember what failed.
Classical planners tend to fail in one of two directions: minimise cost and ignore reliability, or avoid every risk and overpay. Real networks need to tell a fatal hazard from a nuisance — and learn from what already went wrong.
Cat behaviour is the motivating reference: perception, selective obstacle avoidance, detours, memory of outcomes and adaptive exploration. The product model is not limited to cats or to geometric routing.
How it works
Bio-inspired operators. Severity classification, outcome memory, adaptive exploration and alternative generation — conceptually related to PSO, ant-colony and slime-mold methods.
Domain-neutral core. Scenarios, risk models, executors and planners sit behind stable contracts; a GraphBioPlanner runs the same operators over any graph a domain supplies.
Real data. 91k daily rows — 50 SKUs × 5 warehouses × 10 suppliers × 4 regions. Lane reliability is the empirical on-time rate per supplier–warehouse pair, which varies from 0.20 to 1.00.
Controlled ablations. Every operator can be switched off and compared against cost-only, risk-aware and reliability-first classical policies, reporting success rate, cost and risk percentiles, replans and expected economic loss.
What came out of it
The most useful finding was a calibration lesson rather than a headline number.
fewer replans with severity + memory + alternatives, at equal 100% success (3 seeds × 10 episodes)
lower objective cost than the ablated baseline
| One obstacle per lane (50) | 67% |
|---|---|
| One obstacle per supplier (10) | 100% |
With 50 obstacles, hazard exposure stacks up everywhere on the map and compounds over up to 110 hops per route. Aggregating per supplier matches the calibrated density with real severity values unchanged.
| Replans · baseline | 100 |
|---|---|
| Replans · full model | 62 |
| Objective cost · baseline | 100 |
| Objective cost · full model | 89 |
Indexed from the reported ≈38% fewer replans and ≈11% lower objective cost. A regression test locks in the replan advantage.
Severity classification on its own — without alternative generation to recover from a bad choice — performs far worse than the cost-only baseline. Only the full combination outperforms every single-operator ablation.
Stated plainly
- These are algorithm-validation benchmarks, not production evidence. A real claim of superiority would additionally require 100+ scenarios, train/validation/test splits and domain-standard solver baselines.