Fitness Functions Reward What You Wrote, Not What You Meant (Sample)

A genetic algorithm does exactly one thing: it finds whatever maximises the number you gave it. If that number is a poor proxy for what you wanted, the search will find the gap, and it will find it faster than you expect.

The proxy problem#

Suppose you reward distance travelled. The population learns to travel distance. Not to race — to travel distance. If reversing in circles accumulates distance, that is a valid solution to the problem you actually posed.

def fitness(run: Run) -> float:
    # Rewards motion, not progress. Circles score well.
    return run.total_distance

Rewriting it in terms of progress along the track closes that particular gap, and opens a different one.

Things that helped#

• Watch the best individual, not the score. A number going up tells you the search is working. It does not tell you the behaviour is what you wanted.
• Write down what you expect before the run. If the winning strategy surprises you, the fitness function is more likely to be wrong than the agent is to be clever.
• Change one term at a time. Fitness terms interact, and a combined change leaves you unable to attribute the result.

The uncomfortable part#

Most of the design work is not in the algorithm. It is in stating precisely what counts as good, which turns out to be much harder than it sounds — and every vague clause in that definition is somewhere the search will eventually go.