Equal values incorrectly marked as improvements in sampler.py metrics comparison

[Weaverzhu]

https://github.com/codelion/openevolve/blob/main/openevolve/prompt/sampler.py#L263

if prog_value >= parent_value:
    numeric_comparisons_improved.append(True)
else:
    numeric_comparisons_improved.append(False)

if prog_value <= parent_value:
    numeric_comparisons_regressed.append(True)
else:
    numeric_comparisons_regressed.append(False)

When prog_value == parent_value the metrics will be regarded as being improved. When the program makes no difference at all Improvement in all metrics will be prompted to llm, which is very wierd.

[jvm123]

I think this decision might be use case specific and should be configurable.

When using llm_feedback, I see cases where the LLM feedback is not fine-grained enough to rate a new solution as improvement, but the LLM that does the generation indeed did achieve improvements. In essence, I think taking the "newer solution" will more often than not be the right choice in some situations.

In https://github.com/codelion/openevolve/blob/4b099e37c5631a53a9617c367e6acd26b2fe095f/openevolve/database.py#L466 there is a similar case

        # If no metrics, use newest
        if not program1.metrics and not program2.metrics:
            return program1.timestamp > program2.timestamp

        # If only one has metrics, it's better
        if program1.metrics and not program2.metrics:
            return True
        if not program1.metrics and program2.metrics:
            return False

        # Check for combined_score first (this is the preferred metric)
        if "combined_score" in program1.metrics and "combined_score" in program2.metrics:
            return program1.metrics["combined_score"] > program2.metrics["combined_score"]

        # Fallback to average of all numeric metrics
        avg1 = safe_numeric_average(program1.metrics)
        avg2 = safe_numeric_average(program2.metrics)

        return avg1 > avg2

[Weaverzhu]

IMO controling the evolving process and providing metrics feedback to LLM are two different cases. I don't see the reason when the metrics are equal but prompt 'You made some improvements'

[jvm123]

Ok sorry, yes, the imprecise prompt feedback should be improved.

The default behavior when metrics of two programs are equal is yet another topic.

[codelion]

When prog_value == parent_value the metrics will be regarded as being improved. When the program makes no difference at all Improvement in all metrics will be prompted to llm, which is very wierd.

Will fix this to only update when. the value is strictly more.