Evolutionary Algorithm
This library can optimize your agent's performance by way of an evolutionary algorithm, a model of biological evolution: rank fitness, crossover of best performing, mutation, run next generation.
You provide the environment, or "program" as the property is called, (problem to optimze, machine to teach, etc) and a way to calculate and rank fitness, and this algorithm will handle crossover, mutation, and the automatic generation cycles.
It can be implemented on the frontend of any JavaScript web application or on the server itself in a NodeJS application, which of course provides much higher performance. This algorithm is highly computationally intensive and will take time to complete the sufficient generations that will give an optimum solution, but the results are worth it.
To implement the algorithm on the frontend, create an object with the following properties and call these functions:
var input = {
name:"description string",
gens:100, // total generations
runs:20, // runs of scenerio per individual
goal:"max", // "min" and "max" determine internally if the population is ranked by highest or lowest fitness, depending on the goal of the evolution process
pop:100, // total individuals in population
runProcess:"async", // whether the loop that runs each individual in the generation is syncronous or asyncronous
method:"multi-parent", // the crossover method, many parents creating one offspring, or two parents creating many offspring
parents:2, // the number of parents to choose out of the pool of selected individuals
pool:0.1, // the top percentage by fitness of the population to consider as parents for the next generation
splicemin:2, // the minimum length of the dna strand during the splicing procedure
splicemax:12, // the maximum length of the dna strand during the splicing procedure
mutate:0.02 // the percentage of dna that are randomly changed (mutated) before the next generation is run.
// this value is the critical hinge upon which this whole process depends
programInput:programInput, // this is a generic object to attach any data you see fit for your program that may assist in running your program,
// it can be used in any manner you wish
evdelay:0, // delay between the completion of one generation and beginning of the next, 0 for no delay
program:program, // the module for the problem being optimized, details below
}
var evolution = new evolve.module();
evolution.set(input);
evolution.initialize(input);
then start the evolution process like this:
evolution.run(input);
You can stop the evolution process by calling:
evolution.hardStop(input);
Then if you want to restart simply call the run function again with input.gens greater than the current value:
evolution.run(input); //the total generations to be calculateed needs to be increased,
// note that if the goal for the process changes, or the size of the population changes,
// the process will have to be re-initialized and start from the first generation
The program you write that represents the problem being optimized must adhere to the following API:
It must expose the following public functions, each will be called at some point in the evolutionary algorithm process by the evolve package:
{
run:run, // this function will be called to run the evolution process, it may invoke further processes in whatever way it requires to calculate fitness for the individual
gene:gene, // this generates and returns a single new gene, an individual piece of dna
hardStop:hardStop, // this is called to stop immediately any processes running that were started by calling the run function
}
There are a few caveates about the input object for the this package. The program, setEvdata callback, and the completeEvolve callback need to be handled differently depending on whether the evolutionary algorithm and this package run on the frontend or the backend. If the application runs on the frontend, then all three can be included in the input object directly. Obviously, if implementing on the backend, you cannot send module and function objects over an http request. So you will have to devise a way to retrieve your program (already on the backend) when you send this input object and intialize everything. As for the setEvdata callback, this data refers to the best individual of each generation. All that has to be done is to recursively make an http request to an api that calls and returns evolve.module.getBest(). Same for completeEvolve, write a recurive http request that calls and returns evolve.module.running() and when it returns false, simply call whatever function you would have included as your callback in the input object (this will work if the process is stopped automatically or manually).
There are other things to consider for a backend implementation, if there any questions, please don't hesitate to send me an email. When I have the time, I plan on expounding on the wiki and an robust api reference.
My working example of this algorithm put to extensive use can be found at https://evolve.methodswithclass.com. It is applied to various problems like machine learning and image recognition.