Max Planck Institute for Astronomy, European Space Agency.
PaGMO and Astropy - will mainly talk about PaGMO as wrote this and more familiar with this.
PaGMO - Parallel Global Multiobjective Optimiser. Optimisation tool. Initially a trajectory optimisation tool, evolved as a general-purpose optimiser.
Focused on parallel and distributed computing. Can use via C++ or Python.
Astropy (opens new window) - community effort to develop a single core package for Astronomy.
Optimisation - a large area of applied mathematics. “The selection of a best element (with regard to some criteria) from some set of available alternatives.” E.g. Travelling salesman (TSP).
E.g. algorithms: gradient-based methods, evolutionary algorithms, stochastic algorithms.
Interplanetary trajectories - space mission trajectories are defined by sets of parameters: launch date, initial velocity vector, sequence of flybys, sequence of deep-space manoeuvres (DSM).
Usually we want to minimise fuel consumption.
The resulting optimisation problem is hard: multimodal objective function, highly nonlinear, highly dimensional.
Traditionally tackled by teams of human experts.
- genetic algorithms
- differential evolution
- ant-colony optimisation
- artificial bee-colony optimisation
Island model - name inspired by Darwin’s trip to the Galápagos Islands.
History of PaGMO - pattern of scientific programming code: created as part of some research and then abandoned for many years; not useable by anyone else and then picked up later and made more consumable by others.
Initially created Python bindings to initially created C/C++ ‘research’ code (2008-2009) and followed 'eat own dog food’ approach by using it a lot for internal research. Been through 2 GSoCs and now 'fully-fledged’ open source project.
Emphasis on correctness & reproducibility
Powerful driver for innovations in HPC
Code is often written with a one-paper-horizon mindset
Most scientists are not trained in software engineering
The abstract island class includes a problem, an algorithm and a population of candidate solutions, and a virtual evolve() method that dispatches the optimisation (to a thread or a process on another machine)
Implemented via Boost.Python.
serialisation across language boundaries involving virtual classes, base pointers, etc
extension from Python of C++ base classes
working around some of Python’s limitation wrt parallel programming (GIL)
Scientific Python stack: NumPy (crunching results), SciPy (optimisation algorithms), matplotlib, IPython, etc
PaGMO uses: evolution of neural networks for autonomous Martian rovers, selection of Near Earth Asteroids for future human missions.
Handle practical needs of astronomers: units, coordinates, FITS files, cosmological calculations
“One of best community packages ever seen”
Not research package like PaGMO
Heavy reliance on NumPy