Skip to content

Latest commit

 

History

History
33 lines (18 loc) · 3 KB

README.md

File metadata and controls

33 lines (18 loc) · 3 KB

Mantella

Welcome! This mono-repo hosts my research projects on self-developing optimisation algorithms.

Self-developing optimisation methods seek to produce increasingly efficient, domain-specific optimisers and are closely related to (AI) meta-learning techniques.

Based on my professional career as an academic researcher and industrial lead scientist, this repo also includes practical application examples that have stuck with me:

  • Redundancy minimisation of parallel kinematic machines
  • Challenges from the Global Trajectory Optimization Competition
  • Some finance-related topics (based on my current occupation, only as far as compliance permits)

If you want to know more about my work or self-developing optimisation methods in general – like how or why these approaches function – I always have an open ear for you 😊. The easiest way to contact me is via a Github discussion or by email. Alternatively, you may also find me at various scientific conferences throughout the year, like GECCO, ICML or ICCOPT.

Scope

The main purpose of this repo should be to provide an easy and comprehensible access to this field. Instead of complex, extensive libraries, tailored to each use case and every hardware, I provide only the necessary core functions as simple and clear as possible.

To make it easier to transfer these functions to other programming languages/environments and to adapt them to your technical requirements, I will also refrain from external dependencies altogether.

That said, I am happy to list any community project which (re)implements these for others to use. 🚀

History

This project took off in 2013, shortly after I joined the Systems Research and Architecture group at Leibniz Universität Hannover as a research associate. Being paid from the public purse, I wanted to keep my results freely available, following the idea of Open Science.

Back then, the focus was still on automatic algorithm selection and configuration, a precursor to where we are today. Together with wonderful colleagues, students and friends, we explored and developed many different concepts that enriched Mantella every day.

In early 2017, shortly after my exploration of self-developing optimisation algorithms began, things went abruptly quiet. I sailed from academia to industry - which only intensified my work on this topic (albeit less publicly), yet also booked much of my attention. Three years later, I seized the opportunity to rededicate myself to Mantella, revised its purpose, cleaned up a bit and started sharing insights into my work again.

License

The source code is distributed under the MIT license, the content is in the public domain. Feel free to reuse everything as you like.