-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathasynchronous-methods-for-drl.html
1528 lines (952 loc) · 53.4 KB
/
asynchronous-methods-for-drl.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html class="theme-next mist use-motion" lang="zh-Hans">
<head><meta name="generator" content="Hexo 3.8.0">
<meta name="google-site-verification" content="zu-9nWphPjrzXV8v514mkHknIz4dNfHlib56-KNAu44">
<meta charset="UTF-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1">
<meta name="theme-color" content="#222">
<script src="/lib/pace/pace.min.js?v=1.0.2"></script>
<link href="/lib/pace/pace-theme-flash.min.css?v=1.0.2" rel="stylesheet">
<meta http-equiv="Cache-Control" content="no-transform">
<meta http-equiv="Cache-Control" content="no-siteapp">
<script>
(function(i,s,o,g,r,a,m){i["DaoVoiceObject"]=r;i[r]=i[r]||function(){(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;a.charset="utf-8";m.parentNode.insertBefore(a,m)})(window,document,"script",('https:' == document.location.protocol ? 'https:' : 'http:') + "//widget.daovoice.io/widget/356f1943.js","daovoice")
daovoice('init', {
app_id: "356f1943"
});
daovoice('update');
</script>
<link href="/lib/fancybox/source/jquery.fancybox.css?v=2.1.5" rel="stylesheet" type="text/css">
<link href="/lib/font-awesome/css/font-awesome.min.css?v=4.6.2" rel="stylesheet" type="text/css">
<link href="/css/main.css?v=5.1.4" rel="stylesheet" type="text/css">
<link rel="apple-touch-icon" sizes="180x180" href="/images/apple-touch-icon.png?v=5.1.4">
<link rel="icon" type="image/png" sizes="32x32" href="/images/favicon-32x32png?v=5.1.4">
<link rel="icon" type="image/png" sizes="16x16" href="/images/favicon-16x16.png?v=5.1.4">
<link rel="mask-icon" href="/images/logo.svg?v=5.1.4" color="#222">
<meta name="keywords" content="rl,">
<link rel="alternate" href="/atom.xml" title="Keavnn'Blog" type="application/atom+xml">
<script>
(function(){
if(''){
if (prompt('请输入文章密码','') !== ''){
alert('密码错误！');
history.back();
}
}
})();
</script>
<meta name="description" content="本文提出了A3C模型，即Asynchronous Advantage Actor-Critic，是A2C的异步版本，使用CPU多核而不用GPU进行训练，文中说效果比使用GPU反而更好。 推荐： 并行梯度优化的佳作 通俗易懂">
<meta name="keywords" content="rl">
<meta property="og:type" content="article">
<meta property="og:title" content="Asynchronous Methods for Deep Reinforcement Learning">
<meta property="og:url" content="http://StepNeverStop.github.io/asynchronous-methods-for-drl.html">
<meta property="og:site_name" content="Keavnn'Blog">
<meta property="og:description" content="本文提出了A3C模型，即Asynchronous Advantage Actor-Critic，是A2C的异步版本，使用CPU多核而不用GPU进行训练，文中说效果比使用GPU反而更好。 推荐： 并行梯度优化的佳作 通俗易懂">
<meta property="og:locale" content="zh-Hans">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/regression.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/lossfunction.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/gd.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/sgdvsgd.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/a1stepq.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/anstepq.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/a3c.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/table1.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/table2.png">
<meta property="og:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/threeoptimizer.png">
<meta property="og:updated_time" content="2019-05-31T16:01:49.826Z">
<meta name="twitter:card" content="summary">
<meta name="twitter:title" content="Asynchronous Methods for Deep Reinforcement Learning">
<meta name="twitter:description" content="本文提出了A3C模型，即Asynchronous Advantage Actor-Critic，是A2C的异步版本，使用CPU多核而不用GPU进行训练，文中说效果比使用GPU反而更好。 推荐： 并行梯度优化的佳作 通俗易懂">
<meta name="twitter:image" content="http://stepneverstop.github.io/asynchronous-methods-for-drl/regression.png">
<script type="text/javascript" id="hexo.configurations">
var NexT = window.NexT || {};
var CONFIG = {
root: '/',
scheme: 'Mist',
version: '5.1.4',
sidebar: {"position":"left","display":"post","offset":12,"b2t":false,"scrollpercent":true,"onmobile":true},
fancybox: true,
tabs: true,
motion: {"enable":true,"async":true,"transition":{"post_block":"fadeIn","post_header":"slideDownIn","post_body":"slideDownIn","coll_header":"slideLeftIn","sidebar":"slideUpIn"}},
duoshuo: {
userId: '0',
author: '博主'
},
algolia: {
applicationID: '',
apiKey: '',
indexName: '',
hits: {"per_page":10},
labels: {"input_placeholder":"Search for Posts","hits_empty":"We didn't find any results for the search: ${query}","hits_stats":"${hits} results found in ${time} ms"}
}
};
</script>
<link rel="canonical" href="http://StepNeverStop.github.io/asynchronous-methods-for-drl.html">
<title>Asynchronous Methods for Deep Reinforcement Learning | Keavnn'Blog</title>
</head>
<body itemscope="" itemtype="http://schema.org/WebPage" lang="zh-Hans">
<div class="container sidebar-position-left page-post-detail">
<div class="headband"></div>
<a href="https://github.com/StepNeverStop" class="github-corner" aria-label="View source on GitHub" rel="external nofollow" target="_blank"><svg width="80" height="80" viewbox="0 0 250 250" style="fill:#151513; color:#fff; position: absolute; top: 0; border: 0; right: 0;" aria-hidden="true"><path d="M0,0 L115,115 L130,115 L142,142 L250,250 L250,0 Z"/><path d="M128.3,109.0 C113.8,99.7 119.0,89.6 119.0,89.6 C122.0,82.7 120.5,78.6 120.5,78.6 C119.2,72.0 123.4,76.3 123.4,76.3 C127.3,80.9 125.5,87.3 125.5,87.3 C122.9,97.6 130.6,101.9 134.4,103.2" fill="currentColor" style="transform-origin: 130px 106px;" class="octo-arm"/><path d="M115.0,115.0 C114.9,115.1 118.7,116.5 119.8,115.4 L133.7,101.6 C136.9,99.2 139.9,98.4 142.2,98.6 C133.8,88.0 127.5,74.4 143.8,58.0 C148.5,53.4 154.0,51.2 159.7,51.0 C160.3,49.4 163.2,43.6 171.4,40.1 C171.4,40.1 176.1,42.5 178.8,56.2 C183.1,58.6 187.2,61.8 190.9,65.4 C194.5,69.0 197.7,73.2 200.1,77.6 C213.8,80.2 216.3,84.9 216.3,84.9 C212.7,93.1 206.9,96.0 205.4,96.6 C205.1,102.4 203.0,107.8 198.3,112.5 C181.9,128.9 168.3,122.5 157.7,114.1 C157.9,116.9 156.7,120.9 152.7,124.9 L141.0,136.5 C139.8,137.7 141.6,141.9 141.8,141.8 Z" fill="currentColor" class="octo-body"/></svg></a><style>.github-corner:hover .octo-arm{animation:octocat-wave 560ms ease-in-out}@keyframes octocat-wave{0%,100%{transform:rotate(0)}20%,60%{transform:rotate(-25deg)}40%,80%{transform:rotate(10deg)}}@media (max-width:500px){.github-corner:hover .octo-arm{animation:none}.github-corner .octo-arm{animation:octocat-wave 560ms ease-in-out}}</style>
<header id="header" class="header" itemscope="" itemtype="http://schema.org/WPHeader">
<div class="header-inner"><div class="site-brand-wrapper">
<div class="site-meta ">
<div class="custom-logo-site-title">
<a href="/" class="brand" rel="start">
<span class="logo-line-before"><i></i></span>
<span class="site-title">Keavnn'Blog</span>
<span class="logo-line-after"><i></i></span>
</a>
</div>
<h1 class="site-subtitle" itemprop="description">If it is to be, it is up to me.</h1>
</div>
<div class="site-nav-toggle">
<button>
<span class="btn-bar"></span>
<span class="btn-bar"></span>
<span class="btn-bar"></span>
</button>
</div>
</div>
<nav class="site-nav">
<ul id="menu" class="menu">
<li class="menu-item menu-item-home">
<a href="/" rel="section">
<i class="menu-item-icon fa fa-fw fa-home"></i> <br>
首页
</a>
</li>
<li class="menu-item menu-item-about">
<a href="/about/" rel="section">
<i class="menu-item-icon fa fa-fw fa-user"></i> <br>
关于
</a>
</li>
<li class="menu-item menu-item-tags">
<a href="/tags/" rel="section">
<i class="menu-item-icon fa fa-fw fa-tags"></i> <br>
标签
</a>
</li>
<li class="menu-item menu-item-categories">
<a href="/categories/" rel="section">
<i class="menu-item-icon fa fa-fw fa-th"></i> <br>
分类
</a>
</li>
<li class="menu-item menu-item-archives">
<a href="/archives/" rel="section">
<i class="menu-item-icon fa fa-fw fa-archive"></i> <br>
归档
</a>
</li>
<li class="menu-item menu-item-search">
<a href="javascript:;" class="popup-trigger">
<i class="menu-item-icon fa fa-search fa-fw"></i> <br>
搜索
</a>
</li>
</ul>
<div class="site-search">
<div class="popup search-popup local-search-popup">
<div class="local-search-header clearfix">
<span class="search-icon">
<i class="fa fa-search"></i>
</span>
<span class="popup-btn-close">
<i class="fa fa-times-circle"></i>
</span>
<div class="local-search-input-wrapper">
<input autocomplete="off" placeholder="搜索..." spellcheck="false" type="text" id="local-search-input">
</div>
</div>
<div id="local-search-result"></div>
</div>
</div>
</nav>
</div>
</header>
<main id="main" class="main">
<div class="main-inner">
<div class="content-wrap">
<div id="content" class="content">
<div id="posts" class="posts-expand">
<article class="post post-type-normal" itemscope="" itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://StepNeverStop.github.io/asynchronous-methods-for-drl.html">
<span hidden itemprop="author" itemscope="" itemtype="http://schema.org/Person">
<meta itemprop="name" content="Keavnn">
<meta itemprop="description" content="">
<meta itemprop="image" content="/images/Kicon.jpg">
</span>
<span hidden itemprop="publisher" itemscope="" itemtype="http://schema.org/Organization">
<meta itemprop="name" content="Keavnn'Blog">
</span>
<header class="post-header">
<h2 class="post-title" itemprop="name headline">Asynchronous Methods for Deep Reinforcement Learning</h2>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">发表于</span>
<time title="创建于" itemprop="dateCreated datePublished" datetime="2019-05-30T15:22:13+08:00">
2019-05-30
</time>
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-calendar-check-o"></i>
</span>
<span class="post-meta-item-text">更新于:</span>
<time title="更新于" itemprop="dateModified" datetime="2019-06-01T00:01:49+08:00">
2019-06-01
</time>
</span>
<span class="post-category">
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-folder-o"></i>
</span>
<span class="post-meta-item-text">分类于</span>
<span itemprop="about" itemscope="" itemtype="http://schema.org/Thing">
<a href="/categories/ReinforcementLearning/" itemprop="url" rel="index">
<span itemprop="name">ReinforcementLearning</span>
</a>
</span>
</span>
<div class="post-wordcount">
<span class="post-meta-item-icon">
<i class="fa fa-file-word-o"></i>
</span>
<span class="post-meta-item-text">字数统计:</span>
<span title="字数统计">
2.7k
</span>
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-clock-o"></i>
</span>
<span class="post-meta-item-text">阅读时长 ≈</span>
<span title="阅读时长">
10
</span>
</div>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<p>本文提出了A3C模型，即Asynchronous Advantage Actor-Critic，是A2C的异步版本，使用CPU多核而不用GPU进行训练，文中说效果比使用GPU反而更好。</p>
<p>推荐：</p>
<ul>
<li>并行梯度优化的佳作</li>
<li>通俗易懂</li>
</ul>
<a id="more"></a>
<h1 id="简介"><a href="#简介" class="headerlink" title="简介"></a>简介</h1><p>论文地址：<a href="https://arxiv.org/pdf/1602.01783.pdf" rel="external nofollow" target="_blank">https://arxiv.org/pdf/1602.01783.pdf</a></p>
<p>本文提出了一个概念简单、异步梯度优化的轻量级深度强化学习训练框架。提出该框架的初衷是占用少的资源稳定深度神经网络的学习。</p>
<p>该框架可适用于：</p>
<ul>
<li>基于值与基于策略的方法</li>
<li>on-policy与off-policy的方法</li>
<li>离散与连续问题</li>
</ul>
<p>效果：</p>
<ul>
<li>稳定训练</li>
<li>在Atari游戏上使用DQN算法，一个16核CPU比Nvidia K40 GPU快，使用A3C算法可以快一倍</li>
<li>成功适用于很多连续运动学控制问题，例如图像输入的3D迷宫</li>
</ul>
<blockquote>
<p>the sequence of observed data encountered by an online RL agent is non-stationary, and on-line RL updates are strongly correlated. By storing the agent’s data in an experience replay memory, the data can be batched or randomly sampled from different time-steps. Aggregating over memory in this way reduces non-stationarity and decorrelates updates, but at the same time limits the methods to off-policy reinforcement learning algorithms </p>
</blockquote>
<p>文中指出，on-policy方法训练不稳定，数据相关性很强，off-policy机制结合经验池机制减轻了训练的不稳定性和数据相关性。</p>
<p>经验池的缺点：</p>
<ol>
<li>占用内存，增加计算量</li>
<li>需要off-policy算法</li>
</ol>
<p>本文为深度强化学习提供了一个非常不同的范例，不使用经验池机制，而使用异步并行的方法在多个<strong>相同环境</strong>中执行多个智能体。（同时训练多个不同环境没有进行描述和实验）</p>
<p>这种异步并行方式的优点是：</p>
<ol>
<li>实现减轻数据相关性的效果，使训练稳定</li>
<li>可用于大范围on-policy算法，如Sarsa，n-step方法，A-C方法，也可用于off-policy方法，如Q-learning</li>
<li>利用深度神经网络设计算法，保持鲁棒性与有效性（不予置评）</li>
<li>不使用GPU，只使用多核CPU，反而训练时间短</li>
<li>比大规模分布式需要更少的资源占用（内存、算力等）</li>
<li><strong>相比于<a href="https://arxiv.org/pdf/1507.04296.pdf" rel="external nofollow" target="_blank">Gorila</a>异步训练方式，本文中的训练方式不需要中心服务器，其使用的是共享内存模式</strong></li>
<li>单机运行减少了通信（梯度和超参数）的开销</li>
</ol>
<h1 id="文中精要"><a href="#文中精要" class="headerlink" title="文中精要"></a>文中精要</h1><p>由于本文思想非常简单，所以精简描述论文精华</p>
<p>本文使用<strong><a href="https://arxiv.org/pdf/1106.5730.pdf" rel="external nofollow" target="_blank">Hogwild!</a></strong>方式进行异步梯度下降</p>
<p>Hogwild! 这个方法的提出也很偶然，容我步步道来：</p>
<p>我们使用函数对样本集进行拟合如下图所示</p>
<p><img src="./asynchronous-methods-for-drl/regression.png" alt=""></p>
<p>当我们想要判断函数是否拟合的不错，我们往往使用损失函数来衡量，损失越小，则代表函数拟合得越好（但，过拟合不是我们想要的）。</p>
<p><img src="./asynchronous-methods-for-drl/lossfunction.png" alt=""></p>
<p>为了减小损失，我们常用梯度下降算法来优化。标准的梯度下降原理如下图所示，如果函数为凸函数，且更新步长很小，那么在有限步长内总可以下降至函数最小点，即，获得使损失函数最小的参数$\theta^{\ast}$</p>
<p><img src="./asynchronous-methods-for-drl/gd.png" alt=""></p>
<p>之后出现了SGD，也就是随机梯度下降，这种方法差不多在60年代提出，由于思想过于简单，迭代次数很长，一直不被主流优化算法接受。但是，当大数据时代到来时，SGD变成了很普遍的优化方法。</p>
<p>SGD的算法流程如下：</p>
<ul>
<li>选一个初始参数向量$\theta$和正步长$\alpha$</li>
<li>循环直到满足结束条件：<ul>
<li>从训练集中随机选择一个样本$x_{i}$</li>
<li>更新参数$\theta \leftarrow\left(\theta-\alpha \nabla L\left(f_{\theta}\left(x_{i}\right), y_{i}\right)\right)$</li>
</ul>
</li>
</ul>
<p>SGD的优点：</p>
<ol>
<li>少内存占用。SGD不需要所有的样本集进行计算梯度，只需要从样本集中抽取一个样本进行训练。频繁的采样操作可以使用高速缓存来加速训练。</li>
<li>收敛至<strong>可接受的解</strong>速度很快。其实我们不希望看到过拟合，当然也不希望看到欠拟合，SGD正好是这两个极端的trade-off。SGD可以很快的收敛到一个较好的解，相对于样本集较好的解比相对于样本集最好的解的泛化能力可要强得多。下图展示了SGD与标准梯度下降的损失函数曲线比较。</li>
</ol>
<p><img src="./asynchronous-methods-for-drl/sgdvsgd.png" alt=""></p>
<p>看图像可能会觉得SGD并没有严格下降，有时会有损失上升的倾向，但是，总体来看，这种方法最终也是可以收敛到最小值的。总体上，它使损失进行了下降。</p>
<p>对于熟悉并行编程的人来说，如果让他们设计并行随机梯度下降，他们一定会像这样设计：</p>
<ul>
<li>每个线程从训练集随机抽取一个样本$x_{i}$<ul>
<li>锁参数$\theta$</li>
<li>线程读参数$\theta$</li>
<li>线程更新参数$\theta \leftarrow\left(\theta-\alpha \nabla L\left(f_{\theta}\left(x_{i}\right), y_{i}\right)\right)$</li>
<li>解锁</li>
</ul>
</li>
</ul>
<p>更新前锁定参数，更新后解锁参数。对于许多问题，更新这一步骤耗时在微秒级，而锁参数耗时在毫秒级，这意味着锁参数要比更新多占用1000多倍的时间。虽然有一些其他方法可以对该过程进行优化，但差距还是很明显。</p>
<p>如果，<strong>注释掉关于锁的代码呢？</strong>这真是一个大胆的想法，但是Hogwild!就是这么做的。（这都是一个叫Feng Niu的人搞出来的，不知是出于好奇还是在Debug，他在研究加速SGD的时候注释掉了锁的代码，算法不仅有效，还提升了一百多倍。所以说，多试试总是好的。。。）流程如下：</p>
<ul>
<li>每个线程从训练集随机抽取一个样本$x_{i}$<ul>
<li><del>锁参数$\theta $</del></li>
<li>线程读参数$\theta$</li>
<li>线程更新参数$\theta \leftarrow\left(\theta-\alpha \nabla L\left(f_{\theta}\left(x_{i}\right), y_{i}\right)\right)$</li>
<li><del>解锁</del></li>
</ul>
</li>
</ul>
<blockquote>
<p>In a sentence, the main idea of Hogwild! is — “<strong>Remove all thread locks from parallel SGD code</strong>.” In Hogwild!, threads can overwrite each other by writing at the same time and compute gradients using a stale version of the “current solution.” </p>
</blockquote>
<p>我们不禁都有一个疑问，这么做，真的可以吗？还真别说，经过实验表明，该方法取得了多线程的益处却没有数学效率上的负面影响。</p>
<p>本文中的异步更新就是用这种方法，唯一有问题的可能就是两个线程同写，但是即便是同写，写入还是有先后的，最多也就是把前一个线程写入的给覆盖掉，丢弃一个线程的数据更新而已，无伤大雅，至于多个线程写后读、读后写倒都没有大的影响，不影响参数更新与收敛。</p>
<h2 id="伪代码"><a href="#伪代码" class="headerlink" title="伪代码"></a>伪代码</h2><p><strong>本文中所有伪代码都使用Hogwild!方式进行梯度更新。</strong></p>
<p>以下伪代码都是单个线程中actor-learning的操作。每个线程中的actor可以使用不同的探索机制，实验证明不同的探索机制可以提高算法的鲁棒性和性能效果。</p>
<p>首先介绍异步one-step Q-Learning的训练模式</p>
<p><img src="./asynchronous-methods-for-drl/a1stepq.png" alt=""></p>
<p>解析：</p>
<ul>
<li>相比后两个伪代码，该代码中各线程是不需要复制用于选择动作的训练网络（因为各方对target network的定义不同，有些人认为等待赋值的是目标网络，有些人认为需要训练的是目标网络，因此，此处不使用目标网络的术语）的，即每次选择动作，都使用其他线程可能更新过的Q网络进行决策。这是因为不需要使用同一个决策模型向后看多步</li>
<li>$\theta$是动态变化的，即在该线程的训练过程中，其他线程也可能对参数$\theta$进行了更新</li>
<li>$I_{target}$代表双Q学习赋值的间隔</li>
<li>$I_{AsyncUpdate}$代表单线程对共享参数$\theta$更新的间隔</li>
</ul>
<hr>
<p>接下来是异步n-step Q-Learning的训练模式：</p>
<p>相比于n-step，one-step方法中获得的立即奖励$r$只影响导致其产生的$Q(s,a)$，从而通过$Q(s,a)$间接影响其他的动作值，这会使训练过程很慢，因为需要多次更新才能将奖励传播到前面的相关状态和动作。使奖励传播更快的一个方法就是使用n-step回报。</p>
<script type="math/tex; mode=display">
G_{t}=r_{t}+\gamma r_{t+1}+\cdots+\gamma^{n-1} r_{t+n-1}+\max _{a} \gamma^{n} Q\left(s_{t+n}, a\right)</script><p><img src="./asynchronous-methods-for-drl/anstepq.png" alt=""></p>
<p>解析：</p>
<ul>
<li>相比于one-step，该算法为每个线程配置了一个备份网络$\color{red}{\theta’}$</li>
<li>$t_{max}$为n-step中的$n$</li>
</ul>
<hr>
<p>最后是异步A2C，即A3C的训练模式：</p>
<p><img src="./asynchronous-methods-for-drl/a3c.png" alt=""></p>
<p>解析：</p>
<ul>
<li>$t_{max}$为n-step向前看的步数</li>
<li>文中针对该模型将A-C网络架构共享了部分神经网络参数，并且对actor网络的损失函数公式进行了改造：$\nabla_{\theta^{\prime}} \log \pi\left(a_{t} | s_{t} ; \theta^{\prime}\right)\left(R_{t}-V\left(s_{t} ; \theta_{v}\right)\right)+\color{red}{\beta \nabla_{\theta^{\prime}} H\left(\pi\left(s_{t} ; \theta^{\prime}\right)\right)}$，即添加了熵正则化项，它的作用是增加探索，避免网络过早地收敛至局部最优，$\beta$为超参数</li>
</ul>
<p>正态分布的熵可以表示为$-\frac{1}{2}\left(\log \left(2 \pi \sigma^{2}\right)+1\right)$</p>
<h1 id="实验部分"><a href="#实验部分" class="headerlink" title="实验部分"></a>实验部分</h1><p>实验结果视频：</p>
<ul>
<li><a href="https://youtu.be/0xo1Ldx3L5Q" rel="external nofollow" target="_blank">TORCS A3C训练驾驶汽车</a></li>
<li><a href="https://youtu.be/Ajjc08-iPx8" rel="external nofollow" target="_blank">MuJoco 一些训练效果</a></li>
<li><a href="https://youtu.be/nMR5mjCFZCw" rel="external nofollow" target="_blank">Labyrinth 3D迷宫</a></li>
</ul>
<p>Atari 2600 实验结果：</p>
<p><img src="./asynchronous-methods-for-drl/table1.png" alt=""></p>
<p>不同线程数的加速效果：</p>
<p><img src="./asynchronous-methods-for-drl/table2.png" alt=""></p>
<p>文中比较了三种优化函数的性能，分别是：</p>
<ul>
<li>动量SGD，Momentum SGD</li>
<li>RMSProp</li>
<li>Shared RMSProp</li>
</ul>
<p>RMSProp是这样更新的：</p>
<script type="math/tex; mode=display">
g=\alpha g+(1-\alpha) \Delta \theta^{2}</script><script type="math/tex; mode=display">
\theta \leftarrow \theta-\eta \frac{\Delta \theta}{\sqrt{g+\epsilon}}</script><p>$\eta$为学习率，$\alpha$为RMSProp折扣因子</p>
<p>RMSProp与Shared RMSProp的差别就是：</p>
<ul>
<li>Shared RMSProp各线程共享参数$g$，且无锁异步更新</li>
<li>RMSProp各线程独立一个参数$g$</li>
</ul>
<p>实验结果如下：</p>
<p>测试每种算法50次试验，得分从高到低排列，算法为n-step Q-Learning和A3C</p>
<p><strong>综合来看，三种优化方式效果差别不大，但是Shared RMSProp>RMSProp>Momentum SGD</strong></p>
<p><img src="./asynchronous-methods-for-drl/threeoptimizer.png" alt=""></p>
<hr>
<p>实验结果太多，懒得贴了，总之，这种异步框架方法的优点是：</p>
<ul>
<li>使用离线在线策略，基于值基于策略方法，离散连续问题</li>
<li>稳定训练</li>
<li>加速训练</li>
<li>比经验池少消耗资源</li>
</ul>
<h1 id="引用"><a href="#引用" class="headerlink" title="引用"></a>引用</h1><blockquote>
<p><a href="https://medium.com/@krishna_srd/parallel-machine-learning-with-hogwild-f945ad7e48a4" rel="external nofollow" target="_blank">Parallel Machine Learning with Hogwild!</a></p>
</blockquote>
</div>
<div>
<div>
<div style="text-align:center;color: #ccc;font-size:14px;">-------------本文结束<i class="fa fa-heart"></i>感谢您的阅读-------------</div>
</div>
</div>
<div>
<div class="my_post_copyright">
<script src="//cdn.bootcss.com/clipboard.js/1.5.10/clipboard.min.js"></script>
<!-- JS库 sweetalert 可修改路径 -->
<script src="https://cdn.bootcss.com/jquery/2.0.0/jquery.min.js"></script>
<script src="https://unpkg.com/sweetalert/dist/sweetalert.min.js"></script>
<p><span>本文标题:</span><a href="/asynchronous-methods-for-drl.html">Asynchronous Methods for Deep Reinforcement Learning</a></p>
<p><span>文章作者:</span><a href="/" title="访问 Keavnn 的个人博客">Keavnn</a></p>
<p><span>发布时间:</span>2019年05月30日 - 15:05</p>
<p><span>最后更新:</span>2019年06月01日 - 00:06</p>
<p><span>原始链接:</span><a href="/asynchronous-methods-for-drl.html" title="Asynchronous Methods for Deep Reinforcement Learning">http://StepNeverStop.github.io/asynchronous-methods-for-drl.html</a>
<span class="copy-path" title="点击复制文章链接"><i class="fa fa-clipboard" data-clipboard-text="http://StepNeverStop.github.io/asynchronous-methods-for-drl.html" aria-label="复制成功！"></i></span>
</p>
<p><span>许可协议:</span><i class="fa fa-creative-commons"></i> <a rel="external nofollow" href="https://creativecommons.org/licenses/by-nc-sa/4.0/" target="_blank" title="Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)">署名-非商业性使用-相同方式共享 4.0 国际</a> 转载请保留原文链接及作者。</p>
</div>
<script>
var clipboard = new Clipboard('.fa-clipboard');
$(".fa-clipboard").click(function(){
clipboard.on('success', function(){
swal({
title: "",
text: '复制成功',
icon: "success",
showConfirmButton: true
});
});
});
</script>
</div>
<div>
<div style="padding: 10px 0; margin: 20px auto; width: 90%; text-align: center;">
<div>如果您获得了帮助，也可以资助一下小的啦~</div>
<button id="rewardButton" disable="enable" onclick="var qr = document.getElementById('QR'); if (qr.style.display === 'none') {qr.style.display='block';} else {qr.style.display='none'}">
<span>打赏啦</span>
</button>
<div id="QR" style="display: none;">
<div id="wechat" style="display: inline-block">
<img id="wechat_qr" src="/images/wechatpay.jpg" alt="Keavnn 微信">
<p>微信</p>
</div>
<div id="alipay" style="display: inline-block">
<img id="alipay_qr" src="/images/alipay.jpg" alt="Keavnn 支付宝">
<p>支付宝</p>
</div>
</div>
</div>
</div>
<footer class="post-footer">
<div class="post-tags">
<a href="/tags/rl/" rel="tag"> <i class="fa fa-tag"></i> rl</a>
</div>
<div class="post-nav">
<div class="post-nav-next post-nav-item">
<a href="/energy-based-hindsight-experience-prioritization.html" rel="next" title="Energy-Based Hindsight Experience Prioritization">
<i class="fa fa-chevron-left"></i> Energy-Based Hindsight Experience Prioritization
</a>
</div>
<span class="post-nav-divider"></span>
<div class="post-nav-prev post-nav-item">
<a href="/rl2.html" rel="prev" title="RL^2|Fast Reinforcement Learning vis Slow Reinforcement Learning">
RL^2|Fast Reinforcement Learning vis Slow Reinforcement Learning <i class="fa fa-chevron-right"></i>
</a>
</div>
</div>
</footer>
</div>
</article>
<div class="post-spread">
<!-- Go to www.addthis.com/dashboard to customize your tools -->
<div class="addthis_inline_share_toolbox">
<script type="text/javascript" src="//s7.addthis.com/js/300/addthis_widget.js#pubid=ra-5cefbfc88c13b0e7" async="async"></script>
</div>
</div>
</div>
</div>
<div class="comments" id="comments">
<div id="lv-container" data-id="city" data-uid="MTAyMC80MTk0NS8xODQ5MQ=="></div>
</div>
</div>
<div class="sidebar-toggle">
<div class="sidebar-toggle-line-wrap">
<span class="sidebar-toggle-line sidebar-toggle-line-first"></span>
<span class="sidebar-toggle-line sidebar-toggle-line-middle"></span>
<span class="sidebar-toggle-line sidebar-toggle-line-last"></span>
</div>
</div>
<aside id="sidebar" class="sidebar">
<div id="sidebar-dimmer"></div>
<div class="sidebar-inner">
<ul class="sidebar-nav motion-element">
<li class="sidebar-nav-toc sidebar-nav-active" data-target="post-toc-wrap">
文章目录
</li>
<li class="sidebar-nav-overview" data-target="site-overview-wrap">
站点概览
</li>
</ul>
<section class="site-overview-wrap sidebar-panel">
<div class="site-overview">
<div class="site-author motion-element" itemprop="author" itemscope="" itemtype="http://schema.org/Person">
<img class="site-author-image" itemprop="image" src="/images/Kicon.jpg" alt="Keavnn">
<p class="site-author-name" itemprop="name">Keavnn</p>
<p class="site-description motion-element" itemprop="description">If it is to be, it is up to me.</p>
</div>
<nav class="site-state motion-element">
<div class="site-state-item site-state-posts">
<a href="/archives/">
<span class="site-state-item-count">51</span>
<span class="site-state-item-name">日志</span>
</a>
</div>
<div class="site-state-item site-state-categories">
<a href="/categories/index.html">
<span class="site-state-item-count">11</span>
<span class="site-state-item-name">分类</span>
</a>
</div>
<div class="site-state-item site-state-tags">
<a href="/tags/index.html">
<span class="site-state-item-count">26</span>
<span class="site-state-item-name">标签</span>
</a>
</div>
</nav>
<div class="feed-link motion-element">
<a href="/atom.xml" rel="alternate">
<i class="fa fa-rss"></i>
RSS
</a>
</div>
<div class="links-of-author motion-element">
<span class="links-of-author-item">
<a href="https://github.com/StepNeverStop" target="_blank" title="GitHub" rel="external nofollow">
<i class="fa fa-fw fa-github"></i>GitHub</a>
</span>
<span class="links-of-author-item">
<a href="mailto:[email protected]" target="_blank" title="E-Mail" rel="external nofollow">
<i class="fa fa-fw fa-envelope"></i>E-Mail</a>
</span>
</div>
<div class="cc-license motion-element" itemprop="license">
<a href="https://creativecommons.org/licenses/by-nc-sa/4.0/" class="cc-opacity" target="_blank" rel="external nofollow">
<img src="/images/cc-by-nc-sa.svg" alt="Creative Commons">
</a>
</div>
<div class="links-of-blogroll motion-element links-of-blogroll-inline">
<div class="links-of-blogroll-title">
<i class="fa fa-fw fa-link"></i>
推荐阅读
</div>
<ul class="links-of-blogroll-list">
<li class="links-of-blogroll-item">
<a href="https://bluefisher.github.io" title="Fisher Chang" target="_blank" rel="external nofollow">Fisher Chang</a>
</li>
</ul>
</div>
</div>
</section>
<!--noindex-->
<section class="post-toc-wrap motion-element sidebar-panel sidebar-panel-active">
<div class="post-toc">
<div class="post-toc-content"><ol class="nav"><li class="nav-item nav-level-1"><a class="nav-link" href="#简介"><span class="nav-number">1.</span> <span class="nav-text">简介</span></a></li><li class="nav-item nav-level-1"><a class="nav-link" href="#文中精要"><span class="nav-number">2.</span> <span class="nav-text">文中精要</span></a><ol class="nav-child"><li class="nav-item nav-level-2"><a class="nav-link" href="#伪代码"><span class="nav-number">2.1.</span> <span class="nav-text">伪代码</span></a></li></ol></li><li class="nav-item nav-level-1"><a class="nav-link" href="#实验部分"><span class="nav-number">3.</span> <span class="nav-text">实验部分</span></a></li><li class="nav-item nav-level-1"><a class="nav-link" href="#引用"><span class="nav-number">4.</span> <span class="nav-text">引用</span></a></li></ol></div>
</div>
</section>
<!--/noindex-->
</div>
</aside>
</div>
</main>
<footer id="footer" class="footer">
<div class="footer-inner">
<script async src="https://busuanzi.ibruce.info/busuanzi/2.3/busuanzi.pure.mini.js"></script>
<div class="copyright">© <span itemprop="copyrightYear">2020</span>
<span class="with-love">
<i class="fa fa-heart"></i>
</span>
<span class="author" itemprop="copyrightHolder">Keavnn</span>
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-area-chart"></i>
</span>
<span class="post-meta-item-text">Site words total count:</span>
<span title="Site words total count">80.3k</span>
</div>
<div class="powered-by">
<i class="fa fa-user-md"></i><span id="busuanzi_container_site_pv">
本站总访问量<span id="busuanzi_value_site_pv"></span>次
</span>
</div>
<!-- <div class="theme-info">
<div class="powered-by"></div>
<span class="post-count">博客全站共80.3k字</span>
</div> -->