Stochastic Approximation and Brownian Repulsion based Evolutionary Search

Rajdeep Dutta,T Venkatesh Varma,Saikat Sarkar,Mariya Mamajiwala,Noor Awad,Senthilnath Jayavelu,Debasish Roy

2024-12-15

Abstract:Many global optimization algorithms of the memetic variety rely on some form of stochastic search, and yet they often lack a sound probabilistic basis. Without a recourse to the powerful tools of stochastic calculus, treading the fine balance between exploration and exploitation could be tricky. In this work, we propose an evolutionary algorithm (EA) comprising two types of additive updates. The primary update utilizes stochastic approximation to guide a population of randomly initialized particles towards an optimum. We incorporate derivative-free Robbins-Monro type gains in the first update so as to provide a directional guidance to the candidate solutions. The secondary update leverages stochastic conditioning to apply random perturbations for a controlled yet efficient exploration. Specifically, conceptualized on a change of measures, the perturbation strategy discourages two or more trajectories exploring the same region of the search space. Our optimization algorithm, dubbed as SABRES (Stochastic Approximation and Brownian Repulsion based Evolutionary Search), is applied to CEC-2022 benchmark functions on global optimization. Numerical results are indicative of the potentialities of SABRES in solving a variety of challenging multi-modal, non-separable, and asymmetrical benchmark functions.

Optimization and Control

What problem does this paper attempt to address?

### What problems does this paper attempt to solve?

This paper aims to address several key challenges in the field of global optimization. Specifically, it attempts to:

1. **Improve the exploration and exploitation capabilities of existing evolutionary algorithms**:
- Many heuristic - based evolutionary algorithms struggle to strike a balance between exploration and exploitation and are prone to getting trapped in local optimal solutions.
- This paper proposes a new evolutionary algorithm - SABRES (Stochastic Approximation and Brownian Repulsion based Evolutionary Search), which enhances the exploration ability of the algorithm by combining stochastic approximation and Brownian repulsion.

2. **Provide a rigorous probabilistic theoretical foundation**:
- Many existing evolutionary algorithms lack solid probabilistic theoretical support, especially when dealing with high - dimensional, non - convex, and non - smooth functions.
- SABRES utilizes powerful tools in stochastic calculus, especially techniques such as stochastic approximation and Girsanov measure transformation, to provide a rigorous mathematical foundation for the algorithm.

3. **Deal with complex multi - modal, non - separable, and asymmetric benchmark functions**:
- In practical applications, many optimization problems involve multi - modal, non - separable, and asymmetric functions, which place high demands on optimization algorithms.
- This paper demonstrates the application effect of SABRES on the CEC - 2022 benchmark functions, proving its potential in solving these complex problems.

4. **Enhance the robustness and applicability of optimization algorithms**:
- In practical engineering and scientific problems, optimization algorithms need to be able to handle various constraint conditions and have good scalability in different dimensions.
- SABRES improves the robustness of the algorithm by introducing random perturbations and directional guidance, enabling it to better handle nonlinear, non - smooth, and ill - conditioned problems.

### Summary

The core problem of this paper is: how to design an evolutionary algorithm based on a rigorous probabilistic theory to overcome the balance problem between exploration and exploitation in existing algorithms, especially when dealing with complex multi - modal, non - separable, and asymmetric optimization problems. By combining stochastic approximation and Brownian repulsion mechanisms, SABRES provides a brand - new solution and demonstrates superior performance on multiple benchmark functions.