tadiff v0.0.5-d

Tora's Automatic Differentiation

Reverse-mode automatic differentiation for javascript and typescript. Supports symbolic and numeric modes. Only scalars are supported right now.

Demo GitHub

Installation

npm install tadiff

Can also be built by cloning this repository and running npm run build. Tests can be run with npm run test.

Usage

Import the module in typescript import * as tad from "tadiff" or in javascript const tad = require("tadiff");

1. Build an expression

Building a math expression can be done by constructing the tree with the expression classes manually. This can get verbose very quickly so alternatively a string parser is available.

With expression objects:

import * as tad from "tadiff"

const varA = new tad.Variable("a")
const varB = new tad.Variable("b")
const expr = new tad.Divide(new tad.Multiply(new tad.Constant(4), new tad.Exp(new tad.Multiply(varA, varB))), new tad.Abs(new tad.Tan(varB)))

With string parser:

import * as tad from "tadiff"

const expr = tad.parseExpression("4 * exp(a * b) / abs(tan(b))")

// Get the variables of the expression.
// Stored as variables["a"] and variables["b"].
const vars = tad.getAllVariables(expr)

The expression can now be used for numeric evaluation using numbers. Values for all variables have to be passed. The expression can also be printed symbolically.

console.log(expr.evaluateToString())

const evalContext = {
    variableValues: {
        "a": 2,
        "b": 3
    }
}

console.log(expr.evaluate(evalContext))

2. Differentiate an expression

To differentiate an expression we first get all derivatives of an expression using getAllDerivatives. An output derivative has to be passed, this is typically just 1. This will produce all derivatives for all variables within the expression, although each variable has multiple derivatives which have to be summed up using getDerivativeForExpression. Note that this will give back an expression for the derivative. There is nothing special about this expression and it can be used just like any other expression, for example it can be used to build higher order derivatives.

const derivativeA = tad.getDerivativeForExpression(variables["a"], tad.getAllDerivatives(expr, new tad.Constant(1)))
const derivativeB = tad.getDerivativeForExpression(variables["b"], tad.getAllDerivatives(expr, new tad.Constant(1)))

It is also possible to have derivatives within the expression itself.

const derivativeExpr = tad.parseExpression("4 * D(x, cos(x * y))")

3. Interop with mathjs

This library uses mathjs for some of its operations such as parsing and simplification. To convert from tadiff expressions to mathjs expressions use expressionToNode. To convert from mathjs expressions to tadiff expressions use nodeToExpression. Here we will also have to pass a tad.Variable for each of the variable symbols occuring in the node tree.

A good use-case of converting to mathjs nodes is to simplify the tadiff expressions as no simplification is done in tadiff. expressionToNode automatically calls the simplify function.

Expressions

Listed below are the available expressions as both the tadiff classes and the string that can be used for string parsing.

Adding new operations

Creating new operations requires implementing the tad.Expression interface. Most importantly the used inputs are passed through the constructor and need to be returned from getDependencies. evaluateImpl will need to return the numeric result. evaluateToString returns a symbolic string. getDependencyDerivatives needs to return the derivatives for the dependencies in the same order as getDependencies. For more information look at the expressions.ts source code as most expressions are only around 20 lines long. If you do implement a new operation I would be happy to accept a pull request for it.