Feature Suggestion: Add Support for Multi-Objective Optimization Algorithms

[rkhosrowshahi]

First off, thank you for this handy library. It has been a valuable tool in my research and experimentation.

I’d like to propose an enhancement to the library: support for multi-objective optimization (MOO) algorithms, particularly multi-objective evolutionary algorithms (MOEAs).

Currently, evosax is limited to single-objective optimization methods. However, many real-world problems involve multiple conflicting objectives, and solving them requires algorithms that can explore a Pareto-optimal set of solutions.

Adding support for MOO would:

• Address a gap in many general-purpose optimization libraries.
• Attract researchers and practitioners from fields such as multi-task learning, engineering design, neural architecture search, and more.
• Extend the utility of the library beyond single-objective problems, making it more versatile and competitive with libraries like pymoo.

Suggested Additions

Popular MOEAs such as:

1. NSGA-II (for $M\lt3$)
2. NSGA-III (for $M\geq 3$)
3. MOCMAES
4. GDE3
5. MOPSO
6. SPEA2
7. MOEA/D
8. Pareto-front analysis tools (visualization, hypervolume, IGD+, spread metrics, etc.).

I’d be happy to help with implementation or testing if this feature aligns with your development roadmap.

[RobertTLange]

Dear @rkhosrowshahi, please excuse the delayed response. I personally had not planned to add these algorithms. Some of them are already supported by QDAX (e.g. NSGA-2 and SPEA2): https://github.com/adaptive-intelligent-robotics/QDax/blob/main/qdax/baselines/nsga2.py -- I would be happy to welcome these additions though.