Aligner

A little tool for aligning things! This is a demo of tool that could be used to perform physical search space problems where some set of coordinates describe the current alignment of the system and try to find a better alignment given some kind of measure of the goodness of the current alignment.

This demo comes with a Nelder–Mead "optimizer" to get things rolling in the direction of this kind of alignment problem: https://ieeexplore.ieee.org/abstract/document/4026448 but any algorithm can be used.

Note: Since alignment/positioning problems can be pretty domain specific, it's probably better to use the Aligner here as a parent and have your subclass implement introspective logic to match the domain. For example, the 'laser' alignment demo in this readme, may in real life, need to wait a certain dwell time between measurements to allow transient vibrations to damp out sufficiently. This may depend on the step size as well. In that case, a class LaserAligner < Aligner {} would be a great place to overwrite the align method and add those parameters.

Use

A complex system may have many things that need to be aligned. For each one we can make an instance of the Aligner class by choosing the optimization algorithm we'd like to use and the driver.

import { myDriver } from '@my-company/drivers';
import Aligner, { optimizers } from '@kcerb/aligner';

const nelderMead = new optimizers.NelderMead(myDriver.nCoords, {minTolerance: 1e-9, maxIterations: 100});
const laser1Aligner = new Aligner(nelderMead, myDriver);

laser1Aligner.align().then( () => console.log(laser1Aligner.stats));

As long as the driver and optimizer implement the right interfaces, the align call will give the driver a new set of coordinates to try and feed the resultant measurement to the optimizer which will then feed the driver with a new set of coordinates in a loop until the stop condition is reached.

Interfaces

To implement an optimizer or a driver be sure to include the following properties and methods. If this grows in complexity we may want to add a base-class module that provides these as parents to override, or migrate to typescript to get real interfaces.

Optimizer

methods

generator - returns a function generator which gives coordinates of next measurement based on current measurement. Default Aligner will provide two arguments to initialize this generator with: first measurement (number) and first coordinates (number[]).

Driver

properties

currentCoords - returns the current position of the driver as an array of numbers.

methods

measureAt (async) - performs a measurement and returns the value of the measurement which can be fed directly into the optimizer.

Testing

Since drivers are independent, we provide a test driver which can be used to test the performance on sample data of a given algorithm.

import Aligner, { optimizers, drivers } from '@kcerb/aligner';
// test driver will select initial position at random within bounds given by min and max
const testDriver = new drivers.TestDriver('guass-3d', {
    min: [-1, -1, -1],
    max: [1, 1, 1]
});
const nelderMead = new optimizers.NelderMead(testDriver.nCoords, {minTolerance: 1e-9, maxIterations: 100});
const laser1Aligner = new Aligner(nelderMead, testDriver);

laser1Aligner.align().then( () => console.log(laser1Aligner.stats));

ToDo

Next steps:

Add more support for real drivers such as minimum/maximum step-size.
Bloat dependencies a bit for real development (code coverage, linter/style, fancier test suite, actual publication process to npm).
Add visual and other debug tools to help user explore different optimizers effectiveness in aligning a given system.