trial-and-error v1.0.15

Trial and error

This artificial intelligence system has been created with the capability to deduce mathematical operations based on input and output data, and subsequently apply these operations to different sets of data.

It demonstrates exceptional speed when handling straightforward tasks, and its level of complexity adjusts according to the intricacy of the problem it encounters.

The outcome it produces represents the minimal number of operations required to accurately represent the entire training dataset.

Usage

import { train, generate } from "trial-and-error"


const rule = ([a, b, c]) => [a * 5 + c * b, a + c]

let [x,y] = generate(0, 10, 1000, 3, rule, true)

console.time("training")
const model = train(x, y)
console.timeEnd("training") // training: 368.06ms

console.log(model.solutions) // [ 'a*5+b*c', 'a+c' ]

console.log(model.predict([5, 2, 8])) // [ 41, 13 ]
console.log(rule([5, 2, 8])) // [ 41, 13 ]

Here is a example where the solution was now known by me

Rule

Only 0 and 1 if the third number is 0 pass the first two through, if not reverse them

Training data

x = [[1,0,0],[0,1,0],[1,0,1],[0,1,1]]
y = [[1,0],[0,1],[0,1],[1,0]]

Solution

training: 48.458ms
[ 'a+a*c*-2+c', 'a*-1+a*c*2+c*-1+1' ]

Usage

Train Method

[...]
const model = train(x, y, threshold)
[...]

Parameters

x: array of inputs example:

[ [ 8, 1, 9 ], [ 4, 3, 1 ], [ 7, 9, 10 ], [ 11, 9, 7 ], [ 3, 6, 3 ] ]

y: array of outputs example:

[ [ 49 ], [ 23 ], [ 125 ], [ 118 ], [ 33 ] ]

threshold: a positive number, that the closest it is to 0 the more strict the error must be in order to stop looking for a better answer

output

wrapper object with: predict that takes a single input and returns a single output property solutions that returns the solutions in human readable form

generate

[...]
let [x, y] = generate(0, 10, 1000, 3, rule, true)
[...]

parameters

min: how small the number can be

max: how big the number can be

samples: number of samples

sampleLength: length of the samples

rule: rule for creation of samples

cache: if a cache of the result is wanted (this prevents the results from changing in multiple executions)

output

training data on the format accepted by the train function