Experimental Results & Visualization Analysis

In the EvoX multi-objective warehouse robot pathfinding experiment,“getting one nice-looking solution” does not mean the optimization has genuinely made progress. To judge whether a single run is trustworthy, you must obtain two kinds of evidence at the same time: one describes thesearch process(the convergence curves of each objective over generations), and the other describes thefinal solution(the route visualization and metric table under a fixed warehouse layout). The curves only answer “is the search moving at all,” while the route and metrics answer “is the result usable.”

1. Read the Convergence Curves First: What Is Happening in the Search Process

Plot each objective value as a curve against generation; what you observe is thetrend, not the value at any single point. Four common patterns in engineering need to be clearly distinguished:

Continuous improvement: The curve steadily decreases in the early-to-mid stage (or the overall indicator steadily rises), indicating that the population is still searching effectively and that crossover and mutation continue to produce better candidate solutions.
Oscillation: The curve fluctuates up and down without a clear long-term direction, usually caused by evaluation noise, constraint conflicts, or excessive diversity; check the evaluation function and parameter settings.
Stagnation: The curve is nearly flat over a certain interval, but it may still improve again in later generations — this is a “temporary loss of improvement direction.”
Premature convergence: The curve flattens completely at an early stage and never recovers; this is usually due to loss of population diversity or premature compression of the solution space.

2. Next, Look at the Final Route Under the Fixed Warehouse Layout

Overlay the evolved optimal route and the comparison route onthe same fixed warehouse map, and check three things: whether the path is truly reachable, whether obstacle avoidance is reasonable, and whether the final segment near the goal is stable. A fixed layout is a key constraint — only when the environment is unchanged can route differences between different runs or different parameters be attributed to the optimization itself rather than to a changed map.

3. Support Conclusions with Complete Experimental Data

A conclusion only counts if it is reproducible. Please save and verify: the run configuration and random seed, the iteration log, the objective values by generation, the final route, and the evaluation results. When necessary, confirm whether the same conclusion holds consistently through re-runs or cross-checks. If the curve no longer improves markedly in the later stage, use the complete log to determine whether this is premature convergence, evaluation saturation, or simply being close to a usable solution; do not draw conclusions from a single screenshot.

Common Misinterpretations

A curve that smoothly flattens early does not automatically mean “the optimal solution has been found,” nor does it automatically mean “the algorithm has failed”; likewise, a good-looking route does not mean the search is stable. A single best result may be a genuine improvement, but it may also be a product of random fluctuation or evaluation saturation — the process trend and the final result must corroborate each other.

Interpreting Convergence Curves and Pathfinding Results

The figure interprets the search process together with the final solution: the convergence curves reflect each objective's trend over iterations, and you need to distinguish continuous improvement, oscillation, stagnation, and premature convergence; the route map and metric table under the fixed warehouse layout answer whether the result can be used. Only when the process trend and the final route corroborate each other can you rule out misinterpretation caused by a single good-looking result.

Question 1

Regarding the relationship between “convergence curves” and “final route visualization,” which of the following statements is most accurate?

They are unrelated; either one alone is enough to draw a conclusion.

As long as the final route looks reachable, there is no need to look at the convergence curves.

As long as the convergence curve is decreasing, the final route is certainly usable.

The curves reflect whether the search process is progressing, while the route and metric table reflect whether the final solution is usable; the two must be judged together.

Question 2

In a certain EvoX run, the objective values remain flat for a long time after generation 200. What is the most reasonable next step?

Delete this run's record and keep only the one with the best-looking curve.

Use the complete log to determine whether this is premature convergence, evaluation saturation, or simply being close to a usable solution, and re-run with adjusted strategies if necessary.

Change the iteration cap from 500 to 50 to finish the experiment faster.

Simply declare convergence, end the experiment, and write the report.

Question 3

To make the conclusion about “final route performance” reproducible, what should be saved and verified above all?

Save only the candidate route list; no need to save evaluation results.

The run configuration, random seed, iteration log, objective values, final route, and evaluation results.

Save only a screenshot of the final route.

Save only the convergence curve of the last iteration.

Question 4

What is the main purpose of insisting on using “the same fixed warehouse layout” in comparative experiments?

To make the visualization look cleaner and more aesthetically pleasing.

To ensure every run finds the global optimum.

To make route differences between different runs or different algorithms attributable to the optimization itself rather than to environmental changes.

To significantly reduce computation and shorten runtime.

Question 5

What is the most rigorous interpretation of a convergence curve that is “very smooth and stops changing at an early stage”?

You need to check for evaluation saturation or premature loss of population diversity, and cross-validate through re-runs and metrics.

It means the objective function was written incorrectly.

This must be the optimal solution and can be deployed directly.

It means the random seed was set incorrectly.

Case Study: Reproducing Multi-objective Pathfinding Under a Fixed Warehouse Layout

Read the scenario and answer the questions.

You run EvoX on a fixed warehouse layout for multi-objective pathfinding optimization, with path length and turn smoothness as objectives. In one run, the objective curves flatten noticeably after generation 180, producing a seemingly good candidate route. Your team asks you to conclude whether the optimization genuinely progresses and whether the final solution is usable, and to make the conclusion reproducible by others.

1. What data would you collect and verify to support the conclusion? Why can't a single best result alone be used?

Answer:
You need to collect and verify: the run configuration and random seed, the iteration log, the objective values by generation, the final route and evaluation results under the fixed warehouse layout, and the key metric table; re-runs or cross-checks are recommended to confirm the conclusion is stable. A single best result alone cannot be used because a single-point optimum may come from random fluctuation or evaluation saturation — it cannot show whether the search genuinely progressed, cannot be reproduced by others, and cannot determine whether the route is stably usable under the fixed layout.

2. If a re-run produces a similar convergence curve shape but a different final route, how would you judge and handle it?

Answer:
This indicates that the optimization process involves randomness and multiple solutions, so you cannot draw conclusions from a single route. Compare the metric distributions (the means and fluctuation ranges of path length, total turns, and overall cost) rather than single-point values; verify that the seeds and configurations of each run are consistent; and check for premature convergence or evaluation saturation. As remedies, you may increase the number of generations, adjust the diversity-preservation strategy, or fix the evaluation function, and present typical routes and statistical summaries in the report instead of selecting the best-looking route as the sole conclusion.