https://github.com/lancelet/cropduster

Neural Networks for Near-Optimal Control, in *HASKELL*
https://github.com/lancelet/cropduster

crazy haskell neural-network optimal-control

Last synced: 3 months ago
JSON representation

Neural Networks for Near-Optimal Control, in *HASKELL*

• Host: GitHub
• URL: https://github.com/lancelet/cropduster
• Owner: lancelet
• Created: 2023-09-28T11:01:26.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-11-11T00:51:24.000Z (over 1 year ago)
• Last Synced: 2025-01-13T21:27:25.219Z (4 months ago)
• Topics: crazy, haskell, neural-network, optimal-control
• Language: Haskell
• Homepage:
• Size: 107 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Changelog: CHANGELOG.md

Awesome Lists containing this project

README

        
# Optimal Control with Neural Networks

[![CI Build](https://github.com/lancelet/cropduster/actions/workflows/haskell.yml/badge.svg)](https://github.com/lancelet/cropduster/actions/workflows/haskell.yml)

The GitHub Pages site for this repository is at

[https://lancelet.github.io/cropduster](https://lancelet.github.io/cropduster).

The GitHub Pages site currently contains videos of some of the animated

plots generated by the Haskell code.

WIP!

This is a repository containing work for a planned presentation to FP-Syd

(the Sydney, Australia, Functional Programming Group) on near-optimal control

using neural networks.

The presentation will eventually cover:

  - Basics of SGD.

  - Backpropagation.

  - Training supervised networks on batches of examples.

  - Training networks for near-optimal control (ie. control of dynamical systems

    using networks that have been trained by minimising an objective function

    typical of an optimal control problem).

Code examples will (hopefully) include:

  - Least-squares linear fitting by SGD. Compared against the closed-form

    solution for a least-squares fit. This introduces SGD with manual gradient

    calculation to cover the basic approach, in order to motivate automatic

    differentiation.

  - Fitting the parameters of a mass-spring-damper system to an observed

    trajectory. This demonstrates backpropagation through a fixed-step RK4

    ODE solver.

  - Training a network to swing-up a pole in a cart-pole pendulum system.

  - Training a network to do a rocket landing in 2D.

## Plan

Major items

- [x] Show SGD linear fitting.

- [x] Show SGD linear fitting in phase space.

- [x] Implement RK4 for backprop.

- [x] Fit parameters to a mass-spring-damper system using backprop.

- [x] Run compiling and plotting in GitHub CI.

- [x] Trial Haskell chart to see if it has better plotting performance.

- [x] Trial video.js in place of plain HTML video tags.

- [x] Generate slides using Pandoc and reveal.js.

- [x] Clean up `Plot.hs` now I'm using Cairo for plots.

- [x] Set up a flow to render images and then process them to videos straight

      from Haskell to avoid crazy bash scripting.

- [x] Figure out how to include reveal.js presentation in the Hugo output.

- [ ] Tidy spring-damper example; add noise to data; make mass a

      constant since only spring and damping constants are truly free

      parameters.

- [ ] Tidy build.

- [ ] Create other plots using Haskell Chart.

- [ ] Network for cart-pole pendulum balancing example.

- [ ] Network for 2D rocket landing example.

- [ ] Presentation for FP-Syd.

- [ ] Try plyr.js instead of video.js.

## Running

To generate linear fitting example movies:

```

$ ./linfit-examples.sh  # generates movies under the plots directory

```

To generate mass-spring-damper example movies:

```

$ ./msd-examples.sh  # generates movies under the plots directory

```