Allowing Negative Fitness Scores

[alp-c]

Hi Colin,
Is there any specific reason to prevent negative fitness value on https://github.com/colgreen/sharpneat/blob/main/src/SharpNeat/Evaluation/FitnessInfo.cs

This makes harder to defining a scoring system with punishment.

Why This Is Problematic
Disallowing negatives forces us to either:

• Clamp scores at 0, losing valuable information about how bad an agent is.
• Redesign the scoring system (e.g., adding a fixed “buffer” to ensure positivity), which requires knowing the worst possible penalty in advance (often impractical).

[colgreen]

Hi,

The genome selection mechanism in use picks genomes based on their fitness. It's a probabilistic selection based on each genome's fitness as a proportion of Sum(fitness) for all genomes in a species. As such the fitness scores need to be non-negative.

One way around this is to simply do newFitness = Math.Exp(fitness) or some variation on that. This results in a smooth continuous range, with a long tail to to the left asymptotically approaching y=0.