Covariance Matrix Adaptation Study

Covariance Matrix Adaptation (CMA) is a state of the art evolutionary algorithm. In this repository, I explore two applications: portfolio optimization and model merging. This research is part of a broader investigation into evolutionary computation for optimization and model combination. Part of Bocconi University's Computational Modelling course, BSc level.

💰 Evolutionary Portfolio Optimisation

This section is an applied experiment of the CMA strategy for financial Markowitz portfolio optimization. It includes:

• A PDF report detailing the experiment and results
• A Jupyter notebook (.ipynb) file with the code implementation

🧪 Evolutionary Model Merge

This section covers both background information and a research paper that merges the parameters of Large Language Models (LLMs) using CMA, showing how combining fine-tuned models, e.g. math and japanese language, can overlap and combine capabilities, solve math problems formulated in Japanese. Content covered:

1. The Model Merging Problem
2. Useful Background
   • Task Arithmetic
   • TIES-Merging
   • DARE
   • Frankenmerging
   • CMA-ES
3. Evolutionary Model Merge
   • Parameter Space (PS)
   • Data Flow Space (DFS)
   • Interpolation
4. Results
   • Experiments
   • Natural Language Task
   • Computer Vision Task

See PDF presentation for details and references.

🔑License

This project is licensed under the MIT License. See the LICENSE file for details.