在我的代码中，我正在使用张量流制作一个神经网络，并使它学会玩蛇的游戏。
首先，我通过使蛇向10,000个游戏方向随机移动来生成训练数据，并收集每次移动的数据。在我的NN中，我输入： -如果周围有障碍物1个街区 -与食物的角度 -建议的方向
然后产生一个输出，该输出可以是任何实数的数字。
在NN经过训练之后，我通过使其玩游戏来对其进行测试。在这个阶段，我给它所有相同的输入，但是在建议的方向上有所不同。然后，无论哪个建议的方向得分最高，蛇都会朝该方向移动。
我正在模仿这个项目，但是使用javascript： Project
我基本上已经复制了几乎所有特定方面，例如， -NN结构和激活以及优化器和损失 -要提供什么输入和输出
我将提供以下代码。我提供的代码是蛇对象中发生的一切的全部内容。我不包括在内，因为我认为这不是问题。我还将包括我的AI脚本，但我也不认为问题出在该代码中。
问题： 我无法获得我要模仿的项目中看到的结果。蛇也总是死亡。有时，当蛇似乎正朝着食物前进时，它只是在食物周围盘旋而没有实际食用。奇怪的是，我训练的次数越多，情况就越糟。另外，我对时间有疑问。要创建10,000个游戏，我的计算机需要1-2分钟，而训练我的计算机则需要每个时间3-4分钟。考虑我正在使用浏览器和tensorflow.js，这是否花了太多时间？
我认为问题出在我为每个动作分配分数或给定输入中的某个地方。我觉得这些还不够，或者有一些错误，这使这条蛇无法学习我想要的学习方式。
主要游戏循环的代码：
import * as snake from "./snake.js";
import * as AI from "/ai.js";
//Canvas and document information
const c = document.getElementById("can");
const ctx = c.getContext("2d");
const width = can.width;
const height = can.height;
const state = document.getElementById("state");
const score = document.getElementById("score");
const restart1 = document.getElementById("restart");
let model;
let numOfNoSurvive = 0;
let numOfWrongDir= 0;
let numOfRightDir = 0;
//Global constants:
let end = false;
const rez = 20;
//Features to keep track of:
let obsleft
let obsright
let obsfront
//utility functions:
function indexOfMax(arr) {
if (arr.length === 0) {
return -1;
}
var max = arr[0];
var maxIndex = 0;
for (var i = 1; i < arr.length; i++) {
if (arr[i] > max) {
maxIndex = i;
max = arr[i];
}
}
return maxIndex;
}
function randomIntFromInterval(min, max) // min and max included
{
return Math.floor(Math.random() * (max - min + 1) + min);
}
//restart function:
function restart() {
snake1.body = [];
snake1.body[0] = [0, 0];
snake1.dirx = rez;
snake1.diry = 0;
snake1.len = 1;
end = false;
state.innerHTML = "Have Fun!";
score.innerHTML = "0";
food1.x = Math.floor((Math.random() * (width)) / rez) * rez;
food1.y = Math.floor((Math.random() * (height)) / rez) * rez;
}
restart1.onclick = restart;
let dir;
//Input handling
function press(evt) {
if (evt.key == "w") {
if (snake1.diry != rez) {
snake1.diry = -rez;
snake1.dirx = 0;
dir = "u";
}
} else if (evt.key == "s") {
if (snake1.diry != -rez) {
snake1.diry = rez;
snake1.dirx = 0;
dir = "d";
}
} else if (evt.key == "a") {
if (snake1.dirx != rez) {
snake1.diry = 0;
snake1.dirx = -rez;
dir = "l";
}
} else if (evt.key == "d") {
if (snake1.dirx != -rez) {
snake1.diry = 0;
snake1.dirx = rez;
dir = "r";
}
}
}
//document.addEventListener("keydown", press, false);
//Food class:
class Food {
constructor() {
this.x = Math.floor((Math.random() * (width)) / rez) * rez;
this.y = Math.floor((Math.random() * (height)) / rez) * rez;
console.log(this.x, this.y);
this.width = rez;
this.height = rez;
}
update() {
for (let i = 0; i < snake1.body.length; i++) {
if (snake1.body[i][0] == this.x && snake1.body[i][1] == this.y || end == true) {
this.x = Math.floor((Math.random() * (width)) / rez) * rez;
this.y = Math.floor((Math.random() * (height)) / rez) * rez;
snake1.grow();
score.innerHTML = (parseInt(score.innerHTML) + 1).toString();
return true;
}
}
return false;
}
show() {
ctx.fillStyle = "red";
ctx.fillRect(this.x, this.y, this.width, this.height);
}
}
//Creating the objects in the game:
let snake1 = new snake.Snake();
let food1 = new Food();
function update() {
let score;
if (snake1.endGame()) {
state.innerHTML = "Game Over"
end = true;
score = -1;
} else if (!end) {
let head = snake1.body[snake1.body.length - 1];
let hx = head[0];
let hy = head[1];
let fx = food1.x;
let fy = food1.y;
let eat = false;
let prevlen = snake1.len;
let distB = Math.sqrt((Math.pow((fx-hx), 2)) + (Math.pow(fy-hy,2)));
snake1.update();
if(food1.update()){
eat = true;
}
head = snake1.body[snake1.body.length - 1];
hx = head[0];
hy = head[1];
fx = food1.x;
fy = food1.y;
let newlen = snake1.len;
let distA = Math.sqrt((Math.pow((fx-hx), 2)) + (Math.pow(fy-hy,2)));
console.log(hx, hy, fx, fy, distA, distB);
if(distA < distB || newlen > prevlen){
score = 1;
}else{
score = 0;
}
}
return score;
}
function draw() {
if (end) {
ctx.fillStyle = "black";
ctx.fillRect(0, 0, width, height);
} else {
ctx.fillStyle = "white";
ctx.fillRect(0, 0, width, height);
snake1.show();
food1.show();
}
}
function findingLeftRightFront() {
//Get the features and information around the snake
let features = [0, 0, 0];
let head = snake1.body[snake1.body.length - 1];
let hr;
let hl;
let hf;
//set direction ****
if (snake1.dirx == rez && snake1.diry == 0) {
dir = "r";
}
if (snake1.dirx == -rez && snake1.diry == 0) {
dir = "l";
}
if (snake1.dirx == 0 && snake1.diry == rez) {
dir = "d";
}
if (snake1.dirx == 0 && snake1.diry == -rez) {
dir = "u";
}
//Find left right and forward based on direction
switch (dir) {
case "u":
hr = [head[0] + rez, head[1]];
hl = [head[0] - rez, head[1]];
hf = [head[0], head[1] - rez];
break;
case "d":
hr = [head[0] - rez, head[1]];
hl = [head[0] + rez, head[1]];
hf = [head[0], head[1] + rez];
break;
case "l":
hr = [head[0], head[1] - rez];
hl = [head[0], head[1] + rez];
hf = [head[0] - rez, head[1]];
break;
case "r":
hr = [head[0], head[1] + rez];
hl = [head[0], head[1] - rez];
hf = [head[0] + rez, head[1]];
break;
}
//Body part
//right, left, forward
for (let i = 0; i < snake1.body.length - 1; i++) {
let part = snake1.body[i];
if (part[0] == hr[0] && part[1] == hr[1]) {
features[0] = 1;
}
if (part[0] == hl[0] && part[1] == hl[1]) {
features[1] = 1;
}
if (part[0] == hf[0] && part[1] == hf[1]) {
features[2] = 1;
}
}
//Boundaries
if (hr[0] >= width || hr[0] < 0 || hr[1] >= height || hr[1] < 0) {
features[0] = 1;
}
if (hl[0] >= width || hl[0] < 0 || hl[1] >= height || hl[1] < 0) {
features[1] = 1;
}
if (hf[0] >= width || hf[0] < 0 || hf[1] >= height || hf[1] < 0) {
features[2] = 1;
}
return features;
}
function findAngleToFood() {
let head = snake1.body[snake1.body.length - 1];
let hx = head[0];
let hy = head[1];
let fx = food1.x;
let fy = food1.y;
let dirx = snake1.dirx;
let diry = snake1.diry;
//get two vectors a and b: a - unit vector direction of snake, b - unit vector direction from head to food
let a = [(dirx+hx) - hx, (diry+hy) - hy];
let b = [fx - hx, fy - hy];
let dista = Math.sqrt(Math.pow(a[0], 2) + Math.pow(a[1], 2));
let distb = Math.sqrt(Math.pow(b[0], 2) + Math.pow(b[1], 2));
a = [(a[0]/ dista), (a[1]/dista)];
b = [(b[0]/ distb), (b[1]/distb)];
dista = Math.sqrt(Math.pow(a[0], 2) + Math.pow(a[1], 2));
distb = Math.sqrt(Math.pow(b[0], 2) + Math.pow(b[1], 2));
//cos0 = x/y
let x = (a[0] * b[0]) + (a[1] * b[1]);
let y = dista * distb;
let angle = Math.acos(x / y) / (Math.PI);
if(Number.isNaN(angle)){
console.log("CHECK");
angle = 0;
}
console.log(angle);
return angle;
}
function moveSnakeOnValue(val) {
if (val == 0) {
console.log("right");
switch (dir) {
case "u":
snake1.diry = 0;
snake1.dirx = rez;
break;
case "d":
snake1.diry = 0;
snake1.dirx = -rez;
break;
case "l":
snake1.diry = -rez;
snake1.dirx = 0;
break;
case "r":
snake1.diry = rez;
snake1.dirx = 0;
break;
}
}
if (val == 1) {
console.log("left");
switch (dir) {
case "u":
snake1.diry = 0;
snake1.dirx = -rez;
break;
case "d":
snake1.diry = 0;
snake1.dirx = rez;
break;
case "l":
snake1.diry = rez;
snake1.dirx = 0;
break;
case "r":
snake1.diry = -rez;
snake1.dirx = 0;
break;
}
}
if (val == 2) {
console.log("forward");
switch (dir) {
case "u":
snake1.diry = -rez;
snake1.dirx = 0;
break;
case "d":
snake1.diry = rez;
snake1.dirx = 0;
break;
case "l":
snake1.diry = 0;
snake1.dirx = -rez;
break;
case "r":
snake1.diry = 0;
snake1.dirx = rez;
break;
}
}
}
function gameloop() {
if (!end) {
//fix after generating data
let featuresr = findingLeftRightFront();
featuresr.push(findAngleToFood());
featuresr.push(0);
let featuresl = findingLeftRightFront();
featuresl.push(findAngleToFood());
featuresl.push(1);
let featuresf = findingLeftRightFront();
featuresf.push(findAngleToFood());
featuresf.push(2);
let outputr1 = AI.predict(tf.tensor2d([featuresr]));
let outputl1 = AI.predict(tf.tensor2d([featuresl]));
let outputf1 = AI.predict(tf.tensor2d([featuresf]));
let outputr = outputr1.dataSync();
let outputl = outputl1.dataSync();
let outputf = outputf1.dataSync();
let outputs = [outputr[0], outputl[0], outputf[0]];
console.log(featuresr, featuresl, featuresf);
console.log(outputs);
let i = indexOfMax(outputs);
console.log(i);
console.log(snake1.body[0]);
moveSnakeOnValue(i);
//find max from output and set snake direction
let score = update();
draw();
} else {
food1.update();
restart();
}
}
export async function generateTrainingData() {
let dataToTrain = [];
let inputdata = [];
let outputdata = [];
for (let i = 0; i < 10000; i++) {
while (!end) {
//generate input data [right, left, front, angle to food, suggested direction]
let score = update();
let f1 = findingLeftRightFront();
f1.push(findAngleToFood());
let dir = randomIntFromInterval(0, 2);
f1.push(dir);
//0 - right, 1 - left, 2- forward
//move the snake using the suggested direction and get output
moveSnakeOnValue(dir);
if(score == -1){
numOfNoSurvive++;
}else if(score == 0){
numOfWrongDir++;
}else if(score == 1){
numOfRightDir++;
}
//update data
inputdata.push(f1);
outputdata.push([score]);
}
restart();
}
dataToTrain.push(inputdata, outputdata);
console.log(dataToTrain);
return dataToTrain;
}
async function first() {
//order of runtime
//model = await tf.loadLayersModel('/My NN model/my-amazingmodel.json');
let inputs;
let outputs;
await generateTrainingData().then((data) => {
console.log(data[0]);
console.log(data[1]);
inputs = tf.tensor2d(data[0]);
outputs = tf.tensor2d(data[1]);
});
console.log(numOfNoSurvive,numOfWrongDir,numOfRightDir);
console.log(tf.memory().numTensors);
await AI.train(inputs, outputs);
window.setInterval(gameloop, 1000/8);
//await AI.model.save('downloads://my-amazingmodel');
}
first();
AI脚本的代码（使用张量flow.js）：
import * as game from "/game.js";
//Canvas and document information
const c = document.getElementById("can");
const ctx = c.getContext("2d");
const width = can.width;
const height = can.height;
const learningrate = 0.01;
/*
Features:
- The obstacle on the right of the snake head
- The obstacle on the left of the snake head
- The obstacle in front of the snake head
- Suggested direction
So far, these features keep the snake from dying
- The normalized angle between the snake head and the food
This makes the score go up
*/
//Making the model:
export const model = tf.sequential();
model.add(tf.layers.dense({ inputShape: [5], units: 25, activation: 'relu' })); // Hidden
model.add(tf.layers.dense({ units: 1, activation: 'linear' })); // Output
const sgdOpt = tf.train.sgd(learningrate);
const config = {
optimizer: "adam",
loss: "meanSquaredError"
}
model.compile(config);
export async function train(inputs, outputs) {
console.log(inputs.print(), outputs.print());
let history = await model.fit(inputs, outputs, {
batchSize: 32,
epochs: 3,
shuffle: true,
callbacks: {onEpochEnd: (epoch, logs) => console.log("epoch:", epoch, logs.loss)}
});
inputs.dispose();
outputs.dispose();
console.log(history.history.loss[0]);
console.log("Num: ", tf.memory().numTensors);
console.log("training completed");
}
export function predict(input) {
return model.predict(input);
}
我已经为此工作了几天，并感到非常沮丧，因为我似乎无法找到导致蛇AI混乱的错误。
在项目链接中，作者还提供了指向他自己的python代码的链接。我确保所做的一切在概念上都与他的代码相似。但是结果仍然不一样。
我希望这是一个有效的问题，并感谢您阅读。
** edit，我认为如果我添加其余代码，那会更好，因为那里可能有问题。
代码的蛇形部分：
//Canvas and document information
const c = document.getElementById("can");
const ctx = c.getContext("2d");
const width = can.width;
const height = can.height;
const rez = 20;
export class Snake {
constructor() {
this.body = [];
this.body[0] = [0, 0];
this.pressUp = false;
this.pressDown = false;
this.pressLeft = false;
this.pressRight = false;
this.dirx = rez;
this.diry = 0;
this.len = 1;
}
update() {
let head = [...this.body[this.body.length - 1]];
this.body.shift();
head[0] += this.dirx;
head[1] += this.diry;
this.body.push(head);
}
endGame() {
let x = this.body[this.body.length - 1][0];
let y = this.body[this.body.length - 1][1];
if (x >= width || x < 0 || y >= height || y < 0) {
console.log("beep");
return true;
}
for (let i = 0; i < this.body.length - 1; i++) {
let part = this.body[i];
if (part[0] == x && part[1] == y) {
return true;
}
}
return false;
}
show() {
for (let i = 0; i < this.body.length; i++) {
ctx.fillStyle = "black";
ctx.fillRect(this.body[i][0], this.body[i][1], rez, rez);
}
}
grow() {
let head = [...this.body[this.body.length - 1]];
this.len++;
head[0] += this.dirx;
head[1] += this.diry;
this.body.push(head);
}
}
**另一项编辑（抱歉）-如果您想在自己的浏览器中运行它，请执行以下操作： HTML：
<!DOCTYPE html>
<html lang="en" dir="ltr">
<head>
<meta charset="utf-8">
<title>AI</title>
<link rel="stylesheet" type="text/css" href="style.css">
</head>
<body>
<canvas id="can" height="500" width="500" ></canvas>
<button id="restart">Restart</button>
<p id="state">Have Fun!</p>
<p id="score">0</p>
<script src="snake.js" type="module"></script>
<script src="game.js" type="module"></script>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.0.0/dist/tf.min.js"></script>
</body>
</html>
CSS：
body {
margin:0;
}
canvas{
display:block;
border: 3px solid grey;
}
.disable-selection {
-moz-user-select: none; /* Firefox */
-ms-user-select: none; /* Internet Explorer */
-khtml-user-select: none; /* KHTML browsers (e.g. Konqueror) */
-webkit-user-select: none; /* Chrome, Safari, and Opera */
-webkit-touch-callout: none; /* Disable Android and iOS callouts*/
}
问题总结：
如何标记蛇的输入，以便获得良好的训练数据？我在项目中尝试了一个，但是训练得不好。而且，如果您认为我的标签是正确的，那么代码中是否存在错误，无法使蛇正确学习？ <-我在训练和标签方面找不到python代码和javascript之间的任何区别。数字上是否有一些细微调整，以使其学习更好（例如：学习率）。
目标：训练后，snake AI必须能够证明它可以可靠地从10,000场比赛中获得至少10分作为训练数据。该项目中的人获得的收益超过100。
更改和修改：
-同样，它在代码中，但我没有明确说明。我感觉到了我如何奖励行动。到目前为止，模仿该项目，-1尚不存在。 1是朝正确的方向前进。而0的方向错误。这些数字是每个数据的“预期输出”或“标签”
我添加了蛇形部分，如果有问题可能导致代码出现故障
我添加了HTML和CSS，基本上我的所有项目现在都在那里
我再次修复了代码。现在，有关该问题的所有代码均已完全正常运行。输入形状有误。如果您将index.html和所有js文件上传到同一文件夹中。您应该看到它可以正常工作而没有解决问题。 （蛇无法正常学习）。