C# MLAgents代理isn';我在训练方面没有进步
我最近一直在研究一对能够保持自我平衡的腿,这对腿应该可以防止角色摔倒。每个AgentReset,腿部重置其所有必要的因素,如pos、rot和vel,角色下方的地板随机更改旋转角度小于5度。尽管如此,不管我对经纪人的观察次数做了什么,他仍然没有真正从错误中吸取教训。现在,我是机器学习新手,所以对我放松点!我错过了什么?谢谢大家! 一些注释: 我不确定RayPerceptionSensorComponent3D是如何工作的。也许有人能帮我找到正确的方向C# MLAgents代理isn';我在训练方面没有进步,c#,unity3d,machine-learning,artificial-intelligence,ml-agent,C#,Unity3d,Machine Learning,Artificial Intelligence,Ml Agent,我最近一直在研究一对能够保持自我平衡的腿,这对腿应该可以防止角色摔倒。每个AgentReset,腿部重置其所有必要的因素,如pos、rot和vel,角色下方的地板随机更改旋转角度小于5度。尽管如此,不管我对经纪人的观察次数做了什么,他仍然没有真正从错误中吸取教训。现在,我是机器学习新手,所以对我放松点!我错过了什么?谢谢大家! 一些注释: 我不确定RayPerceptionSensorComponent3D是如何工作的。也许有人能帮我找到正确的方向 代理脚本: using MLAgents
代理脚本:
using MLAgents;
using System;
using System.Collections;
using System.Collections.Generic;
using UnityEngine;
using MLAgents.Sensor;
using Random = UnityEngine.Random;
public class BalanceAgent : Agent
{
private BalancingArea area;
public GameObject floor;
public GameObject waist;
public GameObject buttR;
public GameObject buttL;
public GameObject thighR;
public GameObject thighL;
public GameObject legR;
public GameObject legL;
public GameObject footR;
public GameObject footL;
public GameObject[] bodyParts = new GameObject[9];
public HingeJoint[] hingeParts = new HingeJoint[9];
public JointLimits[] jntLimParts = new JointLimits[9];
public Vector3[] posStart = new Vector3[9];
public Vector3[] eulerStart = new Vector3[9];
public RayPerceptionSensorComponent3D raySensors;
float rayDist = 0;
float rayAngle = 0;
Vector3 rayFloorAngle = new Vector3(0,0,0);
Vector3 rayPoint = new Vector3(0,0,0);
int rotAgent = 0;
public void Start() {
bodyParts = new GameObject[] { waist, buttR, buttL, thighR, thighL, legR, legL, footR, footL }; //Waist = 0, footL = 8.
for (int i = 0; i < bodyParts.Length; i++) {
posStart[i] = bodyParts[i].transform.position;
eulerStart[i] = bodyParts[i].transform.eulerAngles;
if (bodyParts[i].GetComponent<HingeJoint>() != null) {
hingeParts[i] = bodyParts[i].GetComponent<HingeJoint>();
jntLimParts[i] = hingeParts[i].limits;
}
}
}
public override void InitializeAgent() {
base.InitializeAgent();
area = GetComponentInParent<BalancingArea>();
}
public override void AgentReset() {
floor.transform.eulerAngles = new Vector3(Random.Range(-5, 5), 0, Random.Range(-5, 5));
print("Reset! - " + rotAgent);
for (int i = 0; i < bodyParts.Length; i++) {
bodyParts[i].transform.position = posStart[i];
bodyParts[i].transform.eulerAngles = eulerStart[i];
if (bodyParts[i].GetComponent<HingeJoint>() != null) {
jntLimParts[i].max = 1;
jntLimParts[i].min = -1;
}
bodyParts[i].GetComponent<Rigidbody>().velocity = Vector3.zero;
bodyParts[i].GetComponent<Rigidbody>().angularVelocity = Vector3.zero;
}
}
public override void AgentAction(float[] vectorAction) {
float buttRDir = 0;
int buttRVec = (int)vectorAction[0];
switch (buttRVec) {
case 1:
buttRDir = 0;
break;
case 2:
buttRDir = .2f;
break;
case 3:
buttRDir = -.2f;
break;
}
jntLimParts[1].max += buttRDir;
jntLimParts[1].min = jntLimParts[1].max - 1;
hingeParts[1].limits = jntLimParts[1];
float buttLDir = 0;
int buttLVec = (int)vectorAction[1];
switch (buttLVec) {
case 1:
buttLDir = 0;
break;
case 2:
buttLDir = .2f;
break;
case 3:
buttLDir = -.2f;
break;
}
jntLimParts[2].max += buttLDir;
jntLimParts[2].min = jntLimParts[2].max - 1;
hingeParts[2].limits = jntLimParts[2];
float thighRDir = 0;
int thighRVec = (int)vectorAction[2];
switch (thighRVec) {
case 1:
thighRDir = 0;
break;
case 2:
thighRDir = .2f;
break;
case 3:
thighRDir = -.2f;
break;
}
jntLimParts[3].max += thighRDir;
jntLimParts[3].min = jntLimParts[3].max - 1;
hingeParts[3].limits = jntLimParts[3];
float thighLDir = 0;
int thighLVec = (int)vectorAction[3];
switch (thighLVec) {
case 1:
thighLDir = 0;
break;
case 2:
thighLDir = .2f;
break;
case 3:
thighLDir = -.2f;
break;
}
jntLimParts[4].max += thighLDir;
jntLimParts[4].min = jntLimParts[4].max - 1;
hingeParts[4].limits = jntLimParts[4];
float legRDir = 0;
int legRVec = (int)vectorAction[4];
switch (legRVec) {
case 1:
legRDir = 0;
break;
case 2:
legRDir = .2f;
break;
case 3:
legRDir = -.2f;
break;
}
jntLimParts[5].max += legRDir;
jntLimParts[5].min = jntLimParts[5].max - 1;
hingeParts[5].limits = jntLimParts[5];
float legLDir = 0;
int legLVec = (int)vectorAction[5];
switch (legLVec) {
case 1:
legLDir = 0;
break;
case 2:
legLDir = .2f;
break;
case 3:
legLDir = -.2f;
break;
}
jntLimParts[6].max += legLDir;
jntLimParts[6].min = jntLimParts[6].max - 1;
hingeParts[6].limits = jntLimParts[6];
float footRDir = 0;
int footRVec = (int)vectorAction[6];
switch (footRVec) {
case 1:
footRDir = 0;
break;
case 2:
footRDir = .2f;
break;
case 3:
footRDir = -.2f;
break;
}
jntLimParts[7].max += footRDir;
jntLimParts[7].min = jntLimParts[7].max - 1;
hingeParts[7].limits = jntLimParts[7];
float footLDir = 0;
int footLVec = (int)vectorAction[7];
switch (footLVec) {
case 1:
footLDir = 0;
break;
case 2:
footLDir = .2f;
break;
case 3:
footLDir = -.2f;
break;
}
jntLimParts[8].max += footLDir;
jntLimParts[8].min = jntLimParts[8].max - 1;
hingeParts[8].limits = jntLimParts[8];
float waistDir = 0;
int waistVec = (int)vectorAction[8];
switch (footLVec) {
case 1:
waistDir = 0;
break;
case 2:
waistDir = .2f;
break;
case 3:
waistDir = -.2f;
break;
}
// waist.transform.Rotate(0, waistDir, 0);
//buttR = vectorAction[0]; //Right or none
//if (buttR == 2) buttR = -1f; //Left
if (waist.transform.position.y > -1.4f) {
AddReward(.02f);
}
else {
AddReward(-.03f);
}
if (waist.transform.position.y <= -3) {
Done();
print("He fell too far...");
}
RaycastHit hit;
Ray r;
if (Physics.Raycast(waist.transform.position, -waist.transform.up, out hit)) {
rayDist = hit.distance;
rayPoint = hit.point;
rayAngle = Vector3.Angle(waist.transform.position, hit.normal);
rayFloorAngle = hit.collider.transform.eulerAngles;
}
}
public override void CollectObservations() {
for(int i = 0; i < bodyParts.Length; i++) {
AddVectorObs(bodyParts[i].transform.position);
AddVectorObs(bodyParts[i].transform.eulerAngles);
AddVectorObs(bodyParts[i].GetComponent<Rigidbody>().velocity);
AddVectorObs(bodyParts[i].GetComponent<Rigidbody>().angularVelocity);
AddVectorObs(jntLimParts[i].max);
AddVectorObs(jntLimParts[i].min);
AddVectorObs(raySensors);
AddVectorObs(rayDist);
AddVectorObs(rayPoint);
AddVectorObs(rayAngle);
AddVectorObs(rayFloorAngle);
}
}
}
using MLAgents;
using System.Collections;
using System.Collections.Generic;
using System.Linq;
using UnityEngine;
public class BalancingArea : Area
{
public List<BalanceAgent> BalanceAgent { get; private set; }
public BalanceAcademy BalanceAcademy { get; private set; }
public GameObject area;
private void Awake() {
BalanceAgent = transform.GetComponentsInChildren<BalanceAgent>().ToList(); //Grabs all agents in area
BalanceAcademy = FindObjectOfType<BalanceAcademy>(); //Grabs balance acedemy
}
private void Start() {
}
public void ResetAgentPosition(BalanceAgent agent) {
//agent.transform.position = new Vector3(area.transform.position.x, 0, area.transform.position.z);
// agent.transform.eulerAngles = new Vector3(0,0,0);
}
// Update is called once per frame
void Update()
{
}
}
using MLAgents;
using System.Collections;
using System.Collections.Generic;
using UnityEngine;
public class BalanceAcademy : Academy
{
}
我认为人们很难快速遵循您的代码。强化学习(rf)有几个主要问题:
- 你用什么样的模型来存储你的知识?流行的有q表,也有使用q学习的深层神经元网络
- 你的奖励功能是什么?你是否使用过任何形式的惩罚
顺便说一句,强化学习与构建在其上的构建块相比非常具有挑战性,就像监督学习方法一样。这也有助于检查它们。我真的可以推荐fastai课程,这些课程都是免费的,而且是我所见过的最重要的机器学习领域的最佳课程。我建议您查看github repo中的ML代理。您可能希望使用与爬虫相同的观察结果: 每个肢体的位置、旋转、速度和角速度加上身体的加速度和角加速度 我还建议从Crawler复制trainer_config.yaml设置,这样您就有了超参数的最佳起点
作为旁注,RayPerceptionSensorComponent3D在某种程度上是独立工作的,您不需要使用AddVectorObs()将其添加到观测中,因为它会在幕后自动执行此操作。您也不再需要为光线感知向空间大小添加任何值。也就是说,对于这个问题,您可能根本不想使用光线感知 为什么不尝试使用transform.positions设置奖励,并在腰部接触地面时结束这一集(如检查腰部和地面之间的碰撞)(或者)检查腰部和其他部位是否处于相同的transform.position并相应地设置奖励
你也可以把你的unity项目文件链接起来。我想看看你在那里做了什么。将此作为一种可能性,但你是否忘记在
命令中添加--train
?这让我有过几次这样的经历。对于207的观测空间,我强烈建议让您的代理训练超过10公里的距离。500万或更多可能是合理的。真的,试着长时间的训练,比如一夜之间,看看是否有什么不同的事情发生。同时拥有ML机器人训练的多个副本也有助于加快训练速度。尝试制作环境的多个副本。(我建议将机器人和飞机放在一个空的游戏对象中,然后复制8-16次)你不需要做任何特殊的事情,只要复制并正常运行,其他副本也会训练。对于第一个问题,很遗憾,我不懂。再说一次,我对这一切都很陌生!我所能说的就是我在用蟒蛇来训练我的ML特工。它只是保存到一个“.nn”文件。尽管对于第二个问题,在靠近底部的代理脚本中,我检查了腰围是否处于某个高度。如果是,+奖励,否则-奖励。谢谢。我一直在尝试设置它,以便让代理知道他什么时候在使用可能的光线投射,但我找不到关于它们的太多信息。我认为光线感知是一种方式,但显然不是。我如何设置光线投射,让我的经纪人看到他是如何与地面成角度的。。?谢谢。RayPerception目前正在实施,它不是为您试图使用它的目的而构建的。您必须修改相当多的ML代理代码,以便在代理的子对象上支持多个RayPerceptionSensorComponents。值得注意的是,光线感知有点像一种廉价的视觉形式。即使人类不能很好地平衡仅有的(超高级)视力,我们也需要一个功能正常的前庭系统。类似地,机器人需要陀螺仪和加速度计。我想你已经选择了最难的方法,哈哈!