Gen AI Library

Gradient descent optimization library for the gen editor's experimental tools.

Overview

The ai library module provides numerical optimization via gradient descent. It offers two variants: a basic gradient descent (go) with ratio-based convergence, and an advanced version (go2) with configurable step functions and epoch limits. Used by experimental gen editor tools for inverse kinematics, body simulation, and parameter fitting.

Functions

Parameters

M.go2(ac, loss, step, epo, s)

Algorithm

Both functions use finite-difference gradient estimation:

1. For each coefficient k, perturb ac[k] by step size
2. Evaluate loss() to estimate gradient direction
3. Update all coefficients proportionally to gradient
4. Repeat until convergence or max iterations

Usage Example

local AI = require('/lua/ge/extensions/editor/gen/lib/ai')

-- Simple optimization
local coeffs = {0.5, 0.3, 0.7}
local function myLoss(ac)
  return (ac[1] - 1)^2 + (ac[2] - 2)^2 + (ac[3] - 3)^2
end
local finalLoss, initialLoss = AI.go(coeffs, myLoss, 0.1)

-- Advanced with custom step
local final, initial = AI.go2(coeffs, myLoss, function(ac, k, s)
  local copy = deepcopy(ac)
  copy[k] = copy[k] + s
  return copy
end, 500, 0.01)

Note on Local Functions

• go(ac, loss, rat) - Local (not exported). Basic gradient descent without custom step function. Used internally.

See Also

• Gen Frame Editor - Uses AI for IK solving
• Gen SolidFlex - Uses AI for physics optimization