Gradient descent vs evolution on three loss landscapes
retained text
I've been getting a bit more into evolutionary algorithms again, so I was testing some loss landscapes where evolution beats vanilla gradient descent (while also trying to make some cool visuals). Round 1, rugged hillside: gradient descent gets stuck in a dip, and evolution reaches the bottom after 750 evaluations. Round 2, smooth slope: gradient descent wins, 108 steps against 570 evaluations. Round 3, flat plateau: the slope is zero, so gradient descent never moves, and evolution reaches the bottom after 840 evaluations. **Edit**: "evolution" here means truncation selection (keep best 30 of 120) plus Gaussian mutation, no crossover.
document fetched 2026-10-03 00:24 · canonical https://reddit.com/r/deeplearning/comments/1wvuu24/gradient_descent_vs_evolution_on_three_loss