examples/Unity ML Agents - 3DBall.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Unity ML Agents - 3DBall\n",
    "\n",
    "In this example, we are going to use a evolutionary algorithm to train a dense neural network for the Unity's [3DBall environment](https://github.com/Unity-Technologies/ml-agents/blob/main/docs/Learning-Environment-Examples.md). The challenge of this environment is to move the agent in a way that doesn't drop the ball."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![Environment](3DBall-Figure1.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Install dependencies and environment\n",
    "\n",
    "First, you need to install the libraries to interact with the Unity environment:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip3 install mlagents==0.27.0\n",
    "!pip3 install gym_unity"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For the environment, there are two options:\n",
    "\n",
    "- Use the pre-built version (Windows): [3DBall Windows Build](https://drive.google.com/file/d/150UDRhKwpFWaKiehBYE3Qk7xfLE2GGTN/view?usp=sharing), download and extract.\n",
    "- Build manually:\n",
    "  - Install Unity and clone the ML Agents repository: [Installation Instructions](https://github.com/Unity-Technologies/ml-agents/blob/main/docs/Installation.md)\n",
    "  - Open the repository project in Unity (\"Project\" folder)\n",
    "  - Open the \"3DBall\" scene, and keep only one copy of the environment (delete \"3DBall (1-11)\", keeping only the first \"3DBall\"), as the interface only works with scenes with an agent.\n",
    "  - Build the scene: [Using an Environment Executable\n",
    "](https://github.com/Unity-Technologies/ml-agents/blob/main/docs/Learning-Environment-Executable.md), follow the instructions in \"Building the 3DBall environment\"\n",
    "\n",
    "Once the environment is built, change the path in the block below to the environment path:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "args = {\"env_path\": \"D:\\\\Github\\\\ml-agents\\\\builds\\\\3dball_single\\\\UnityEnvironment.exe\"}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluator\n",
    "\n",
    "Our agent will be a dense neural network, which will have the weights encoded by the individual. The class that will be our fitness function is defined in the file [\"unity3dball_evaluator.py\"](unity3dball_evaluator.py)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "import evolvepy as ep\n",
    "import numpy as np\n",
    "from matplotlib import pyplot as plt\n",
    "import unity3dball_evaluator as unity3dball"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We create the evaluator:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluator = ep.evaluator.ProcessEvaluator(unity3dball.Unity3DBallEvaluator, args = args)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And the descriptor of our individual. Let's use a network with 2 hidden layers, each with 20 units.:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import nn\n",
    "descriptor = nn.create_descriptor(input_size=8, output_size=2, units=[20,20])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Generator\n",
    "\n",
    "let's start by defining our population size. As we are going to parallelize the evaluations, it is interesting to use a size multiple of the number of processors, using all the computational power:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "96"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import multiprocessing as mp\n",
    "\n",
    "population_size = (100//mp.cpu_count())*mp.cpu_count()\n",
    "population_size"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Our generator will have elitism, keeping the best 5% individuals of the generation, and random predation, exchanging the worst 75% individuals for better individuals (in the other 25%):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "first = ep.generator.Layer()\n",
    "sort = ep.generator.Sort()\n",
    "\n",
    "first.next = sort\n",
    "\n",
    "concat = ep.generator.Concatenate()\n",
    "\n",
    "predation = ep.generator.RandomPredation(int(0.75*population_size))\n",
    "combine = ep.generator.CombineLayer(ep.generator.selection.tournament, ep.generator.crossover.one_point)\n",
    "mutation = ep.generator.mutation.NumericMutationLayer(ep.generator.mutation.sum_mutation, 1.0, 0.9, (-0.5, 0.5))\n",
    "filter0 = ep.generator.FilterFirsts(int(np.floor(0.95*population_size)))\n",
    "\n",
    "sort.next = predation\n",
    "predation.next = combine\n",
    "combine.next = mutation\n",
    "mutation.next = filter0\n",
    "filter0.next = concat\n",
    "\n",
    "filter1 = ep.generator.FilterFirsts(int(np.ceil(0.05*population_size)))\n",
    "\n",
    "sort.next = filter1\n",
    "filter1.next = concat\n",
    "\n",
    "generator = ep.generator.Generator(first_layer=first, last_layer=concat, descriptor=descriptor)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evolver\n",
    "\n",
    "Finally, we define our Evolver. We gonna use dynamic mutation to avoid getting stuck at a local maximum:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "dyn_mut = ep.callbacks.DynamicMutation([mutation.name], refinement_patience=5, exploration_patience=5, refinement_steps=5)\n",
    "\n",
    "evolver = ep.Evolver(generator, evaluator, population_size, [dyn_mut])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluation and results\n",
    "\n",
    "Let's evolve for 70 generations:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "hist, last_pop = evolver.evolve(70)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And see the results:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(hist.max(axis=1))\n",
    "plt.xlabel(\"Generation\")\n",
    "plt.ylabel(\"Fitness\")\n",
    "plt.title(\"Evolution History\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can see that the agent manages to improve in the environment, reaching the maximum possible fitness of 100:\n",
    "\n",
    "![](3DBall-Figure2.jpg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's see how the best individual fared in the environment:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "best = last_pop[np.argmax(hist[-1])]\n",
    "\n",
    "test_evaluator = unity3dball.Unity3DBallEvaluator(show=True, args = args)\n",
    "test_evaluator.evaluate([best])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We have a recorded video of the best individual performed in the test (click on the picture to see the video):"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[![Video](3DBall-Figure1.png)](https://youtu.be/iCOS1ag6cPk)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, by adding a WandbCallback to the evolver, we can generate a more detailed log of the evolution: [example log](https://wandb.ai/eltoncn/EvolvePy%20Example/runs/2xuoq82x?workspace=user-eltoncn)"
   ]
  }
 ],
 "metadata": {
  "interpreter": {
   "hash": "2db524e06e9f5f4ffedc911c917cb75e12dbc923643829bf417064a77eb14d37"
  },
  "kernelspec": {
   "display_name": "Python 3.8.5 64-bit",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.5"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}