Simulated Isotropic Annealing with Normal-Distribution Search Steps

Summary

This article modifies simulated annealing by replacing uniformly distributed search steps with normally distributed steps centered on the current solution. The proposed Simulated Isotropic Annealing (SIA) method generates normal variates, limits their range, and scales them to the allowed parameter interval. The rationale is to concentrate proposals near the current point while retaining some chance of larger moves. The article situates this change within simulated annealing’s broader use of randomness, acceptance of worse solutions, and gradual cooling, while noting the difficulty of setting temperature and cooling parameters and the challenge of large search spaces.

The author tests the approach on Rastrigin, Forest, and Megacity benchmark functions at several dimensions, reporting results for a fixed number of function evaluations. In those tests, the normal-step version performs worse than the uniform-step baseline, and increasing the spread parameter further worsens results. Later claims of broad efficiency and resistance to local trapping should be read alongside these benchmark-specific findings. These are optimization experiments, not trading backtests, so they do not establish improved trading performance or universal superiority.

Key ideas

• SIA changes simulated annealing proposals from uniform steps to normally distributed steps around the current solution.
• The method limits and rescales proposals to fit the selected parameter range.
• The article’s benchmark tests report worse results than the uniform-step baseline for the tested normal distribution.
• Search quality depends on exploration behavior and configuration, especially in high-dimensional spaces.
• Benchmark function results alone do not establish benefits for live trading or strategy optimization.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.