Genetic Algorithms and Design of Experiments

Tuning Evolutionary Algorithms for UK Logistics Route Optimisation

This dissertation explored how Design of Experiments can be used to tune a genetic algorithm for solving route optimisation problems in UK logistics.

The work focused on the Traveling Salesman Problem, where a vehicle must visit a set of locations and return to its starting point while minimising total distance.

In logistics, this is not just an academic problem. Poor route planning increases mileage, fuel usage, emissions, vehicle utilisation, driver time, and operational cost.

The central research question was:

Can a genetic algorithm, tuned using a structured experimental method, outperform manual transport planning in a UK logistics context?

Abstract

This dissertation investigated the use of Design of Experiments to tune the crossover and mutation rates of a genetic algorithm designed to solve logistics route optimisation problems.

The study was grounded in the UK road logistics sector, where many routing decisions are still made manually. Although commercial optimisation tools exist, they can be expensive, difficult to trust, or poorly aligned to the realities of smaller and medium-sized logistics operations.

A genetic algorithm was developed as a practical optimisation artefact. Design of Experiments was then used to assess how changes in crossover rate, mutation rate, and route size affected the quality of the routes produced.

The research showed that a tuned genetic algorithm could outperform original manual planning in the sample data. It also demonstrated that parameter tuning matters, that crossover plays a significant role in genetic algorithm performance, and that optimal parameter settings are context-dependent.

Research Objectives

The dissertation had six main objectives:

Develop a genetic algorithm capable of solving Travelling Salesman Problem style route optimisation problems.
Investigate how genetic algorithm parameter values influence performance.
Understand whether parameter values interact with each other.
Provide guidance on suitable parameter settings for UK logistics route problems.
Develop a structured methodology for tuning genetic algorithms using Design of Experiments.
Test whether a genetic algorithm can outperform manual transport planning.

Research Philosophy

The philosophy behind the dissertation was pragmatic and applied.

Rather than treating optimisation as a purely mathematical exercise, the study started from a real operational problem: logistics planners need good routes quickly, and perfect theoretical optimisation is often less useful than timely, explainable, operationally acceptable improvement.

The work sat between four areas:

Computer science, because it involved building a working genetic algorithm.
Operations research, because it addressed route optimisation and decision quality.
Empirical experimentation, because algorithm performance was tested through measurable outcomes.
Practical logistics, because the data and problem context came from real-world UK transport operations.

The core idea was that algorithmic performance is context-dependent.

A genetic algorithm that works well in one context may not work well in another unless its parameters are tuned to the problem domain. This was important because many optimisation techniques are tested against benchmark datasets, while real logistics work involves operational constraints, imperfect data, planner judgement, and commercial trade-offs.

mindmap
  root((Research Philosophy))
    Pragmatic
      Useful over perfect
      Operational value matters
      Real logistics data
    Empirical
      Test assumptions
      Measure performance
      Compare outcomes
    Context-dependent
      No universal best setting
      Route size matters
      Algorithm behaviour changes
    Product-oriented
      Trust matters
      Adoption matters
      Decision support over black box automation

Background

The Travelling Salesman Problem

The Travelling Salesman Problem is a classic optimisation problem. A vehicle must visit each location once, return to its starting point, and minimise the total distance travelled.

The problem becomes difficult because the number of possible route sequences grows extremely quickly as the number of stops increases.

For small routes, a planner may be able to reason through a few options. As the number of stops grows, the number of possible routes becomes too large to evaluate manually.

flowchart LR
    A[More Stops] --> B[More Possible Route Sequences]
    B --> C[Harder Manual Planning]
    C --> D[Greater Risk of Suboptimal Routes]
    D --> E[Extra Distance]
    E --> F[Higher Cost and Emissions]

Classical Route Planning Limitations

Manual and traditional planning approaches can struggle with:

Combinatorial explosion as route size increases.
Limited ability to evaluate all possible route sequences.
Planner time pressure.
Inconsistent route quality.
Increased mileage, cost, and emissions.
Limited trust in black-box optimisation systems.

Genetic Algorithm Potential

Genetic algorithms offer a practical alternative because they:

Work well with combinatorial optimisation problems.
Explore a population of possible solutions rather than a single route.
Can produce useful near-optimal answers without checking every possible route.
Can be tuned to suit the specific problem context.
Are flexible enough to support different fitness functions and constraints.

How the Genetic Algorithm Works

Genetic algorithms are inspired by biological evolution. They generate a population of possible solutions, evaluate them, select better-performing solutions, combine them through crossover, and introduce controlled variation through mutation.

For route optimisation, each individual solution represents a possible sequence of locations.

flowchart TD
    A[Initial Route Population] --> B[Evaluate Fitness]
    B --> C{Stop Criteria Met?}
    C -- No --> D[Select Better Routes]
    D --> E[Crossover]
    E --> F[Mutation]
    F --> G[New Route Population]
    G --> B
    C -- Yes --> H[Best Route Found]

Methodology

The dissertation developed a genetic algorithm artefact and then used Design of Experiments to test how different parameter settings affected performance.

The main experimental factors were:

Crossover rate
Mutation rate
Problem size, based on the number of stops in the route

The main response variable was route quality, measured by the distance achieved by the genetic algorithm compared with the original manually planned route.

flowchart TD
    A[Build Genetic Algorithm Artefact] --> B[Collect UK Logistics Route Data]
    B --> C[Define Experimental Factors]
    C --> D[Crossover Rate]
    C --> E[Mutation Rate]
    C --> F[Problem Size]

    D --> G[Run DOE Experiments]
    E --> G
    F --> G

    G --> H[Measure Route Performance]
    H --> I[Compare Against Manual Planning]
    I --> J[Analyse Results]
    J --> K[Draw Conclusions]

Design of Experiments Approach

Design of Experiments was used because genetic algorithm tuning can easily become guesswork.

Changing one parameter at a time can miss interactions between variables. For example, the best mutation rate may depend on the crossover rate, route size, or other algorithm design choices.

DOE provided a more structured way to test parameter settings and understand whether those settings had a meaningful effect on performance.

flowchart LR
    A[Algorithm Parameters] --> B[Experimental Design]
    B --> C[Controlled Runs]
    C --> D[Measured Route Quality]
    D --> E[Statistical Analysis]
    E --> F[Parameter Insight]
    F --> G[Better Tuning Decisions]

Technical Design Decisions

The dissertation made several important design decisions.

Permutation Encoding

The genetic algorithm used permutation encoding, because route sequencing is naturally represented as an ordered list of locations.

This reduced the risk of generating invalid routes and made the representation more suitable for the Travelling Salesman Problem.

Focused Parameter Testing

The study focused on crossover and mutation rates rather than attempting to tune every possible genetic algorithm variable.

This kept the research controlled and allowed the dissertation to focus on whether DOE could be used effectively as a tuning method.

Real Logistics Data

The algorithm was tested using logistics data rather than only academic benchmark datasets.

This was important because the dissertation was interested in operational usefulness, not just theoretical performance.

flowchart TD
    A[Design Decision] --> B[Use Permutation Encoding]
    A --> C[Use Real Logistics Data]
    A --> D[Focus on Crossover and Mutation]
    A --> E[Avoid Excessive Domain Assumptions]

    B --> F[Valid Route Sequences]
    C --> G[Operational Relevance]
    D --> H[Clear Experimental Focus]
    E --> I[Less Overfitted Artefact]

Key Findings

1. Genetic algorithms can outperform manual route planning

The developed genetic algorithm produced route improvements compared with the original manual planning.

This supports the idea that accessible optimisation methods can generate real operational value in logistics.

2. Tuning matters

The research showed that parameter settings affected performance.

This is one of the most important conclusions. The value is not simply in having an algorithm. The value is in ensuring that the algorithm is tuned for the problem it is solving.

3. Design of Experiments is a useful tuning method

DOE provided a disciplined method for testing genetic algorithm settings.

It offered a more structured alternative to trial-and-error tuning and helped identify the effect of parameter choices on algorithm performance.

4. Crossover was significant

The study supported the view that crossover is a major source of the genetic algorithm’s ability to create better solutions.

Crossover helped combine useful route characteristics from existing solutions into improved new routes.

5. Mutation was not significant in this context

Mutation did not show a significant effect in the study.

This was an important finding because it contradicted some previous research. It suggested that mutation may be less important in this specific UK logistics TSP context, or that its role depends on other unmodelled variables.

6. Problem size changes the best settings

The dissertation found that ideal settings varied depending on route size.

This reinforced the argument that genetic algorithm tuning is context-sensitive.

7. Other variables probably matter

Although the genetic algorithm performed well, the regression models had limited explanatory power.

This suggested that crossover and mutation were not the only factors influencing performance. Other algorithmic or route-specific variables may also have played an important role.

flowchart TD
    A[Research Findings] --> B[GA Improved on Manual Planning]
    A --> C[Tuning Affected Performance]
    A --> D[DOE Was Useful]
    A --> E[Crossover Was Significant]
    A --> F[Mutation Was Not Significant]
    A --> G[Problem Size Changed Best Settings]
    A --> H[Other Variables Likely Matter]

    B --> I[Operational Value]
    C --> J[Do Not Treat Algorithms as Plug-and-Play]
    D --> K[Structured Tuning Beats Guesswork]
    E --> L[Crossover Drives Route Improvement]
    F --> M[Mutation Effect Was Limited in This Context]
    G --> N[Context-sensitive Configuration]
    H --> O[Further Research Required]

Product Management Interpretation

The dissertation has a clear product lesson.

The value of optimisation software is not simply that it contains an algorithm. The value comes from how well the algorithm fits the operational problem, how explainable the output is, and whether users can trust its recommendations.

In transport planning, users are unlikely to accept a black-box answer just because the software says it is optimal. They need confidence that the recommendation reflects the real operating context.

This makes optimisation a product problem as much as a mathematical problem.

flowchart LR
    A[Algorithm] --> B[Correctness]
    A --> C[Performance]
    A --> D[Explainability]
    A --> E[Trust]

    B --> F[Useful Product]
    C --> F
    D --> F
    E --> F

    F --> G[Adoption by Transport Planners]

Practical Application

The dissertation showed that a tuned genetic algorithm can improve on manual transport planning in a UK logistics context. The wider product lesson is that optimisation tools need more than mathematical strength: they need tuning, context, explainability, and user trust.

Conclusions

The dissertation concluded that a tuned genetic algorithm can provide practical value for UK logistics route optimisation.

The work showed that algorithm design and parameter tuning should be treated as empirical activities.

It is not enough to select an algorithm from the literature and assume it will perform well. The algorithm must be tested against the operating environment it is intended to support.

The key conclusion was:

Optimisation performance is not only about the algorithm itself; it is about the fit between the algorithm, its parameters, and the problem context.

For logistics operations, this means route optimisation tools should be developed and evaluated using realistic transport data, operational constraints, and measurable business outcomes.

flowchart TD
    A[Algorithm Choice] --> D[Optimisation Performance]
    B[Parameter Tuning] --> D
    C[Problem Context] --> D
    D --> E[Route Quality]
    E --> F[Operational Benefit]
    F --> G[Lower Mileage]
    F --> H[Lower Cost]
    F --> I[Lower Emissions]
    F --> J[Better Planning Support]

Limitations

The study had several limitations.

It focused on a subset of UK logistics route problems, so the findings should not be assumed to apply to every routing scenario.

Only selected genetic algorithm parameters were tested, mainly crossover and mutation rates.

Other factors may also influence performance, including:

Population size
Selection method
Stopping criteria
Encoding strategy
Fitness function design
Local search methods
Route complexity
Data quality

The research also focused on the Travelling Salesman Problem rather than the broader Vehicle Routing Problem.

Real-world transport planning often includes additional constraints such as vehicle capacity, driver hours, delivery windows, customer priority, vehicle type, depot restrictions, and route balancing.

Future Research

Future research could extend the work by testing a wider range of algorithm parameters and route types.

Useful extensions would include:

Comparing genetic algorithms against other optimisation methods.
Testing larger and more varied logistics datasets.
Extending the model from TSP to full Vehicle Routing Problems.
Introducing operational constraints such as time windows, capacity, and driver hours.
Exploring adaptive genetic algorithms where parameters change during the run.
Investigating hybrid approaches that combine genetic algorithms with local search.

flowchart TD
    A[Future Research] --> B[Wider Parameter Testing]
    A --> C[More Logistics Data]
    A --> D[Compare Other Algorithms]
    A --> E[Move from TSP to VRP]
    A --> F[Add Real-world Constraints]
    A --> G[Explore Adaptive Tuning]
    A --> H[Hybrid Optimisation Methods]

    E --> I[Capacity]
    E --> J[Time Windows]
    E --> K[Driver Hours]
    E --> L[Vehicle Types]

Reflection

This dissertation was ultimately about more than route optimisation.

It demonstrated the value of combining technical implementation with structured experimentation.

It also showed the importance of understanding the business context before applying advanced algorithms.

The work remains relevant because many organisations still face the same problem: complex operational decisions are often made manually, supported by experience rather than systematic optimisation.

The dissertation showed that even where perfect optimisation is unrealistic, better decision support is achievable.

The practical lesson is simple:

Good algorithms need good context, good tuning, and good evaluation.

That principle continues to shape how I think about product development, operational systems, and the use of AI and optimisation in real-world software.