Flutter遗传进化神经网络插件genetically_evolving_neural_network的应用

Flutter遗传进化神经网络插件genetically_evolving_neural_network的应用

1. 插件概述

genetically_evolving_neural_network 是一个Flutter插件，它模拟了基因进化的过程，通过交叉繁殖和基因突变来优化神经网络。每个实体（Entity）由一个神经网络组成，其“DNA”可以视为神经网络的结构，而“基因”则是神经网络中的感知器（Perceptrons）。通过评估每个实体的表现并选择表现较好的实体进行繁殖，插件实现了进化的过程。

2. 工作原理

• 初始种群生成：插件会生成一个随机的初始种群，每个实体包含一个随机生成的神经网络。
• 适应度评分：每个神经网络都会根据自定义的 FitnessService 进行评估，获得一个适应度评分（fitness score），评分越高表示该神经网络的表现越好。
• 繁殖与变异：表现较好的实体有更高的概率被选为父母，生成下一代实体。在繁殖过程中，父母的基因会被组合，并有一定概率发生突变。
• 进化循环：这个过程会不断重复，逐渐优化种群中的神经网络。

3. 使用步骤

3.1 定义 FitnessService

FitnessService 是用来评估每个实体表现的关键部分。你需要根据具体的任务定义一个 FitnessService，例如判断一个数字是正数还是负数、预测下一步棋的走法等。

/// The inputs for this Neural Network, from -1 to 1 in increments of 0.1.
List<double> get inputs => List.generate(10, (index) => index * 0.1)
  ..addAll(List.generate(9, (index) => (index + 1) * -0.1));

/// The scoring function that will be used to evolve entities of a population
class PositiveNumberFitnessService extends GENNFitnessService {
  @override
  Future<double> gennScoringFunction({
    required GENNNeuralNetwork neuralNetwork,
  }) async {
    // Calculate how many correct guesses were made
    return inputs.fold(0, (previousValue, input) {
      final guess = neuralNetwork.guess(inputs: [input])[0];
      // Only add a point if the neural network guesses correctly
      if ((input > 0 && guess > 0) || (input <= 0 && guess == 0)) {
        return previousValue + 1;
      }
      return previousValue;
    }).toDouble();
  }
}

3.2 配置 GeneticEvolutionConfig

GeneticEvolutionConfig 用于定义种群的大小、突变率等参数。你可以根据任务的需求调整这些参数。

final config = GENNGeneticEvolutionConfig(
  numInitialInputs: 1, // 1 input into neural network
  numOutputs: 1, // 1 output from neural network
  layerMutationRate: 0.1, // 10% chance to add/remove layer to network
  perceptronMutationRate: 0.2, // 20% chance to add/remove perceptron to layer
  mutationRate: 0.05, // 5% chance to mutate an existing perceptron
);

3.3 创建 GENN 对象

使用 FitnessService 和 GeneticEvolutionConfig 创建 GENN 对象，并开始进化过程。

final genn = GENN.create(
  config: config,
  fitnessService: PositiveNumberFitnessService(),
);

// 获取下一代实体
final nextGen = await genn.nextGeneration();

4. 示例代码

以下是一个完整的示例，展示了如何使用 genetically_evolving_neural_network 插件来构建一个简单的正负数分类器。

import 'package:flutter/material.dart';
import 'package:genetically_evolving_neural_network/genetically_evolving_neural_network.dart';

/// The inputs for this Neural Network, from -1 to 1 in increments of 0.1.
List<double> get inputs => List.generate(10, (index) => index * 0.1)
  ..addAll(List.generate(9, (index) => (index + 1) * -0.1));

/// The scoring function that will be used to evolve entities of a population
class PositiveNumberFitnessService extends GENNFitnessService {
  @override
  Future<double> gennScoringFunction({
    required GENNNeuralNetwork neuralNetwork,
  }) async {
    // Calculate how many correct guesses were made
    return inputs.fold(0, (previousValue, input) {
      final guess = neuralNetwork.guess(inputs: [input])[0];
      // Only add a point if the neural network guesses correctly
      if ((input > 0 && guess > 0) || (input <= 0 && guess == 0)) {
        return previousValue + 1;
      }
      return previousValue;
    }).toDouble();
  }
}

void main() {
  runApp(const MyApp());
}

class MyApp extends StatefulWidget {
  const MyApp({super.key});

  @override
  State<MyApp> createState() => _MyAppState();
}

class _MyAppState extends State<MyApp> {
  /// Represents the FitnessService used to drive this GENN example.
  final GENNFitnessService fitnessService = PositiveNumberFitnessService();

  /// The current generation of Neural Networks.
  GENNGeneration? generation;

  /// The Genetically Evolving Neural Network object.
  late final GENN genn;

  @override
  void initState() {
    // Declare a config with specific mutation rates.
    final config = GENNGeneticEvolutionConfig(
      populationSize: 20,
      numOutputs: 1,
      mutationRate: 0.1,
      numInitialInputs: 1,
      layerMutationRate: 0.25,
      perceptronMutationRate: 0.4,
    );

    // Create the GENN object from the incoming config and fitness service.
    genn = GENN.create(
      config: config,
      fitnessService: fitnessService,
    );

    // Initialize the first generation
    genn.nextGeneration().then((value) {
      setState(() {
        generation = value;
      });
    });

    super.initState();
  }

  @override
  Widget build(BuildContext context) {
    final generation = this.generation;
    if (generation == null) {
      return const CircularProgressIndicator();
    }

    return MaterialApp(
      title: 'GENN Example',
      home: Scaffold(
        body: Center(
          child: Column(
            mainAxisAlignment: MainAxisAlignment.center,
            children: [
              Text('代数: ${generation.wave.toString()}'),
              Text('最高得分: '
                  '${generation.population.topScoringEntity.fitnessScore} '
                  '总分: ${inputs.length}.01'),
              Text('最高得分实体的层数: '
                  '${generation.population.topScoringEntity.maxLayerNum + 1}'),
              Text('最高得分实体的感知器数量: '
                  '${generation.population.topScoringEntity.dna.genes.length}'),
            ],
          ),
        ),
        floatingActionButton: FloatingActionButton.extended(
          label: const Text('下一世代'),
          onPressed: () {
            // Set the next Generation to be displayed
            genn.nextGeneration().then((value) {
              setState(() {
                this.generation = value;
              });
            });
          },
        ),
      ),
    );
  }
}

5. 更复杂的示例

除了简单的正负数分类器，genetically_evolving_neural_network 还可以用于解决更复杂的问题，例如：

• 逻辑异或（Logical XOR）：判断输入是否满足逻辑异或条件，即只有一个输入为1.0，其他输入为0.0时返回正确结果。
• 图像数字分类器：识别像素化图像所代表的数字（0-9）。

5.1 逻辑异或示例

/// This fitness service will be used to score a logical XOR calculator. The
/// Neural Network should only be rewarded for guessing "yes" when there is a
/// single input of 1.0 and both other inputs are 0.
class LogicalXORGENNVisualizationFitnessService extends GENNFitnessService {
  @override
  Future<double> gennScoringFunction({
    required GENNNeuralNetwork neuralNetwork,
  }) async {
    // Collect all the guesses from this NeuralNetwork
    final guesses = getNeuralNetworkGuesses(neuralNetwork: neuralNetwork);

    // Declare a variable to store the sum of all errors
    var errorSum = 0.0;

    // Cycle through each guess to check its validity
    for (int i = 0; i < guesses.length; i++) {
      // Calculate the error from this guess
      final error = (targetOutputsList[i] == guesses[i]) ? 0 : 1;

      // Add this error to the errorSum
      errorSum += error;
    }

    // Calculate the difference between a perfect score (8) and the total
    // errors. A perfect score would mean zero errors with 8 correct answers,
    // meaning a perfect score would be 8.
    final diff = inputList.length - errorSum;

    // To make the better performing Entities stand out more in this population,
    // use the following equation to calculate the FitnessScore.
    //
    // 4 to the power of diff
    return pow(4, diff).toDouble();
  }
}

5.2 图像数字分类器示例

/// This class will be used to score a Number Classifier in tandem with a
/// Neural Network.
class NumberClassifierFitnessService extends GENNFitnessService {
  @override
  Future<double> gennScoringFunction({
    required GENNNeuralNetwork neuralNetwork,
  }) async {
    // Collect all the guesses from this NeuralNetwork
    final guesses = getNeuralNetworkGuesses(neuralNetwork: neuralNetwork);

    // Declare a variable to store the sum of points scored
    int points = 0;

    // Cycle through each guess to check its validity
    for (int i = 0; i < guesses.length; i++) {
      final NaturalNumber targetOutput = targetOutputsList[i];
      final NaturalNumber guessOutput = guesses[i];

      if (targetOutput == guessOutput) {
        // Guessing correctly will give you a point.
        points++;
      }
    }

    // To make the better performing Entities stand out more in this population,
    // use the following equation to calculate the FitnessScore.
    //
    // 4 to the power of points
    return pow(4, points).toDouble();
  }
}

6. 文件解析

你还可以将特定的世代写入文件，并从文件中读取特定的世代。

• 写入文件：

genn.writeGenerationToFile();

• 从文件加载：

genn.loadGenerationFromFile(wave: 10);