{"id":42817,"date":"2026-06-24T17:02:55","date_gmt":"2026-06-24T14:02:55","guid":{"rendered":"https:\/\/fatihsoysal.com\/blog\/?p=42817"},"modified":"2026-06-24T17:04:24","modified_gmt":"2026-06-24T14:04:24","slug":"derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme","status":"publish","type":"post","link":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/","title":{"rendered":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme"},"content":{"rendered":"<h2>Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme<\/h2>\n<p>Derin Peki\u015ftirmeli \u00d6\u011frenme (Deep Reinforcement Learning &#8211; DRL), bilgisayar sistemlerinin karma\u015f\u0131k ortamlarda insan benzeri veya \u00fcst\u00fcn performans sergilemesini sa\u011flayan g\u00fc\u00e7l\u00fc bir yapay zeka alan\u0131d\u0131r. Atari oyunlar\u0131 gibi sim\u00fclasyon ortamlar\u0131nda botlar geli\u015ftirmek, DRL&#8217;nin pratik uygulamalar\u0131n\u0131n en bilinen \u00f6rneklerindendir. Ancak, DRL modellerinin e\u011fitimi s\u0131ras\u0131nda kar\u015f\u0131la\u015f\u0131lan en temel zorluklardan biri, modelin hem \u00f6\u011frenme veri setine a\u015f\u0131r\u0131 uyum sa\u011flamas\u0131n\u0131 (y\u00fcksek sapma) hem de k\u00fc\u00e7\u00fck veri de\u011fi\u015fikliklerine kar\u015f\u0131 a\u015f\u0131r\u0131 hassas olmas\u0131n\u0131 (y\u00fcksek varyans) \u00f6nlemektir. Bu makalede, derin peki\u015ftirmeli \u00f6\u011frenme ba\u011flam\u0131nda sapma-varyans dengesinin \u00f6nemini inceleyecek, bu dengeyi anlamak i\u00e7in temel kavramlar\u0131 a\u00e7\u0131klayacak ve OpenAI Gym k\u00fct\u00fcphanesini kullanarak bir Atari botu in\u015fa etme s\u00fcrecinde bu prensipleri nas\u0131l uygulayabilece\u011fimizi detayland\u0131raca\u011f\u0131z.<\/p>\n<h3>Sapma ve Varyans Kavramlar\u0131<\/h3>\n<p>Makine \u00f6\u011frenmesinde, bir modelin genelleme yetene\u011fi, yani daha \u00f6nce g\u00f6rmedi\u011fi veriler \u00fczerindeki performans\u0131, sapma ve varyans aras\u0131ndaki dengeye ba\u011fl\u0131d\u0131r. Bu iki kavram, modelin tahminlerinin ger\u00e7ek de\u011ferden ne kadar sapma g\u00f6sterdi\u011fini a\u00e7\u0131klamak i\u00e7in kullan\u0131l\u0131r.<\/p>\n<h4>Sapma (Bias)<\/h4>\n<p>Sapma, bir modelin tahminlerinin ger\u00e7ek de\u011ferden ortalama olarak ne kadar uzak oldu\u011funu \u00f6l\u00e7er. Y\u00fcksek sapma, modelin ger\u00e7ek veriyi yeterince iyi yakalayamad\u0131\u011f\u0131n\u0131 g\u00f6sterir. Bu durum, genellikle modelin \u00e7ok basit olmas\u0131ndan veya yeterince karma\u015f\u0131k \u00f6r\u00fcnt\u00fcleri \u00f6\u011frenememesinden kaynaklan\u0131r. Bir DRL ajan\u0131 ba\u011flam\u0131nda, y\u00fcksek sapma, ajan\u0131n oyunun temel mekaniklerini anlamad\u0131\u011f\u0131n\u0131, optimal stratejileri \u00f6\u011frenemedi\u011fini veya \u00e7evrenin \u00f6nemli \u00f6zelliklerini g\u00f6z ard\u0131 etti\u011fini g\u00f6sterebilir. Basit bir model, karma\u015f\u0131k bir oyunun t\u00fcm n\u00fcanslar\u0131n\u0131 \u00f6\u011frenmekte yetersiz kalabilir.<\/p>\n<h4>Varyans (Variance)<\/h4>\n<p>Varyans, bir modelin tahminlerinin, e\u011fitim verisindeki k\u00fc\u00e7\u00fck de\u011fi\u015fikliklere ne kadar duyarl\u0131 oldu\u011funu \u00f6l\u00e7er. Y\u00fcksek varyans, modelin e\u011fitim verisine a\u015f\u0131r\u0131 uyum sa\u011flad\u0131\u011f\u0131n\u0131 (overfitting) g\u00f6sterir. Bu, modelin e\u011fitim verisindeki g\u00fcr\u00fclt\u00fcy\u00fc veya rastgeleli\u011fi de \u00f6\u011frendi\u011fi anlam\u0131na gelir. Sonu\u00e7 olarak, model yeni, g\u00f6r\u00fclmemi\u015f veriler \u00fczerinde k\u00f6t\u00fc performans g\u00f6sterir \u00e7\u00fcnk\u00fc e\u011fitim verisindeki spesifik \u00f6rneklere a\u015f\u0131r\u0131 derecede adapte olmu\u015ftur. DRL&#8217;de y\u00fcksek varyans, ajan\u0131n belirli bir oyun oturumunda veya belirli bir rastgele ba\u015flang\u0131\u00e7 durumunda \u00e7ok iyi performans g\u00f6sterip, k\u00fc\u00e7\u00fck bir de\u011fi\u015fiklikte tamamen \u00e7\u00f6kebilece\u011fi anlam\u0131na gelebilir. Bu, ajan\u0131n oyunun genel stratejisini de\u011fil, sadece belirli durumlar\u0131 ezberledi\u011fini g\u00f6sterir.<\/p>\n<h4>Sapma-Varyans Dengesi<\/h4>\n<p>\u0130deal bir model, hem d\u00fc\u015f\u00fck sapmaya hem de d\u00fc\u015f\u00fck varyansa sahip olmal\u0131d\u0131r. Ancak pratikte, sapma ve varyans aras\u0131nda bir \u00f6d\u00fcnle\u015fme (trade-off) vard\u0131r. Genellikle, sapmay\u0131 azaltmaya y\u00f6nelik bir \u00e7aba (daha karma\u015f\u0131k bir model kullanarak), varyans\u0131 art\u0131rma e\u011filimindedir ve tersi de ge\u00e7erlidir. DRL&#8217;de bu dengeyi kurmak, ajan\u0131n oyunun kurallar\u0131n\u0131 ve stratejilerini genelleme yetene\u011fini en \u00fcst d\u00fczeye \u00e7\u0131kar\u0131rken, a\u015f\u0131r\u0131 uyumu \u00f6nlemek i\u00e7in kritik \u00f6neme sahiptir.<\/p>\n<h3>Derin Peki\u015ftirmeli \u00d6\u011frenme ve Sapma-Varyans<\/h3>\n<p>DRL modelleri, derin sinir a\u011flar\u0131n\u0131 peki\u015ftirmeli \u00f6\u011frenme algoritmalar\u0131yla birle\u015ftirir. Bu derin sinir a\u011flar\u0131, karma\u015f\u0131k, y\u00fcksek boyutlu girdileri (\u00f6rne\u011fin, oyun ekran g\u00f6r\u00fcnt\u00fcleri) i\u015fleyebilir ve eylemler i\u00e7in de\u011fer fonksiyonlar\u0131n\u0131 veya politikalar\u0131 tahmin edebilir. Ancak, derin sinir a\u011flar\u0131n\u0131n do\u011fas\u0131 gere\u011fi, sapma ve varyans sorunlar\u0131 daha da karma\u015f\u0131k hale gelebilir.<\/p>\n<h4>Derin Sinir A\u011flar\u0131n\u0131n Rol\u00fc<\/h4>\n<p>Derin sinir a\u011flar\u0131, b\u00fcy\u00fck miktarda parametreye sahip karma\u015f\u0131k fonksiyon yakla\u015f\u0131klar\u0131d\u0131r. Bu karma\u015f\u0131kl\u0131k, a\u011flar\u0131n potansiyel olarak \u00e7ok d\u00fc\u015f\u00fck sapmaya sahip olmas\u0131n\u0131 sa\u011flar, yani karma\u015f\u0131k veri \u00f6r\u00fcnt\u00fclerini \u00f6\u011frenebilirler. Ancak, ayn\u0131 zamanda a\u011flar\u0131n e\u011fitim verisine a\u015f\u0131r\u0131 uyum sa\u011flama (y\u00fcksek varyans) riskini de art\u0131r\u0131r. \u00d6zellikle, yeterli veri olmad\u0131\u011f\u0131nda veya e\u011fitim s\u00fcreci iyi y\u00f6netilmedi\u011finde bu risk daha da b\u00fcy\u00fcr.<\/p>\n<h4>Peki\u015ftirmeli \u00d6\u011frenme Ba\u011flam\u0131nda Zorluklar<\/h4>\n<p>DRL&#8217;de sapma ve varyans, geleneksel denetimli \u00f6\u011frenmeden farkl\u0131 zorluklar sunar:<\/p>\n<p>*   <strong>Veri Toplama S\u00fcreci:<\/strong> DRL&#8217;de veri, ajan taraf\u0131ndan ortamla etkile\u015fim kurularak toplan\u0131r. Bu veri toplama s\u00fcreci do\u011fas\u0131 gere\u011fi rastgele ve g\u00fcr\u00fclt\u00fcl\u00fcd\u00fcr. Ajan\u0131n politikas\u0131 geli\u015ftik\u00e7e, toplad\u0131\u011f\u0131 verinin da\u011f\u0131l\u0131m\u0131 da de\u011fi\u015fir. Bu, varyans\u0131n kontrol\u00fcn\u00fc zorla\u015ft\u0131r\u0131r.<br \/>\n*   <strong>Hedeflerin De\u011fi\u015fkenli\u011fi:<\/strong> Peki\u015ftirmeli \u00f6\u011frenmede, hedef (\u00f6d\u00fcl sinyali) zamanla de\u011fi\u015febilir ve gelecekteki \u00f6d\u00fclleri tahmin etmek karma\u015f\u0131kt\u0131r. Bu belirsizlik, modelin sapmas\u0131n\u0131 art\u0131rabilir.<br \/>\n*   <strong>\u00d6\u011frenme Algoritmalar\u0131:<\/strong> DRL algoritmalar\u0131 (\u00f6rn. Q-learning, Policy Gradients) genellikle yinelemeli olarak \u00e7al\u0131\u015f\u0131r ve \u00f6nceki tahminlere dayan\u0131r. Bu, hata yay\u0131l\u0131m\u0131na ve sapma\/varyans sorunlar\u0131n\u0131n birikmesine yol a\u00e7abilir.<br \/>\n*   <strong>Hiperparametre Ayar\u0131:<\/strong> DRL modelleri, \u00f6\u011frenme oran\u0131, indirim fakt\u00f6r\u00fc, a\u011f mimarisi ve d\u00fczenlile\u015ftirme teknikleri gibi bir\u00e7ok hiperparametreye sahiptir. Bu hiperparametrelerin yanl\u0131\u015f ayarlanmas\u0131, sapma veya varyans sorunlar\u0131na neden olabilir.<\/p>\n<h3>OpenAI Gym ile Atari Botu \u0130n\u015fa Etme<\/h3>\n<p>OpenAI Gym, peki\u015ftirmeli \u00f6\u011frenme algoritmalar\u0131n\u0131 geli\u015ftirmek ve kar\u015f\u0131la\u015ft\u0131rmak i\u00e7in standart bir aray\u00fcz sunan pop\u00fcler bir k\u00fct\u00fcphanedir. Atari oyunlar\u0131, Gym&#8217;de pop\u00fcler bir ortam grubudur \u00e7\u00fcnk\u00fc y\u00fcksek boyutlu g\u00f6rsel girdiye sahiptirler ve karma\u015f\u0131k karar verme s\u00fcre\u00e7leri gerektirirler.<\/p>\n<h4>Temel Bile\u015fenler<\/h4>\n<p>Bir Atari botu in\u015fa etmek i\u00e7in genellikle \u015fu bile\u015fenlere ihtiya\u00e7 duyar\u0131z:<\/p>\n<p>1.  <strong>Ortam (Environment):<\/strong> OpenAI Gym&#8217;den bir Atari oyunu (\u00f6rn. <code>Pong-v0<\/code>, <code>Breakout-v0<\/code>). Ortam, durumlar\u0131 (g\u00f6r\u00fcnt\u00fcler) sa\u011flar, eylemleri kabul eder ve \u00f6d\u00fclleri ile yeni durumlar\u0131 d\u00f6nd\u00fcr\u00fcr.<br \/>\n2.  <strong>Ajan (Agent):<\/strong> Karar verme mant\u0131\u011f\u0131n\u0131 i\u00e7eren k\u0131s\u0131m. Bu, genellikle bir derin sinir a\u011f\u0131 taraf\u0131ndan temsil edilen bir politik veya de\u011fer fonksiyonunu \u00f6\u011frenen bir DRL algoritmas\u0131d\u0131r.<br \/>\n3.  <strong>Sinir A\u011f\u0131 Mimarisi:<\/strong> Girdi (oyun ekran\u0131) ve \u00e7\u0131kt\u0131 (eylemler i\u00e7in olas\u0131l\u0131klar veya Q-de\u011ferleri) aras\u0131nda haritalama yapan derin sinir a\u011f\u0131. Atari oyunlar\u0131 i\u00e7in genellikle Evri\u015fimsel Sinir A\u011flar\u0131 (Convolutional Neural Networks &#8211; CNN&#8217;ler) kullan\u0131l\u0131r.<br \/>\n4.  <strong>Peki\u015ftirmeli \u00d6\u011frenme Algoritmas\u0131:<\/strong> Ajan\u0131n deneyimlerinden \u00f6\u011frenmesini sa\u011flayan algoritma (\u00f6rn. DQN, A2C, PPO).<\/p>\n<h4>Sapma-Varyans Dengelemesi \u0130\u00e7in Teknikler<\/h4>\n<p>OpenAI Gym ile bir Atari botu geli\u015ftirirken sapma-varyans dengesini y\u00f6netmek i\u00e7in \u00e7e\u015fitli teknikler kullan\u0131labilir:<\/p>\n<h5>1. Model Karma\u015f\u0131kl\u0131\u011f\u0131n\u0131n Y\u00f6netimi (Sapmay\u0131 Azaltma, Varyans\u0131 Kontrol Alt\u0131nda Tutma)<\/h5>\n<p>*   <strong>Sinir A\u011f\u0131 Mimarisi:<\/strong> CNN&#8217;lerin katman say\u0131s\u0131, filtre boyutu ve aktivasyon fonksiyonlar\u0131 gibi mimari se\u00e7imleri modelin karma\u015f\u0131kl\u0131\u011f\u0131n\u0131 etkiler. \u00c7ok s\u0131\u011f bir a\u011f y\u00fcksek sapmaya yol a\u00e7abilirken, \u00e7ok derin ve geni\u015f bir a\u011f a\u015f\u0131r\u0131 uyum riskini art\u0131r\u0131r. Deneyerek en uygun mimariyi bulmak \u00f6nemlidir.<br \/>\n*   <strong>\u00d6zellik \u00c7\u0131kar\u0131m\u0131:<\/strong> Oyun ekranlar\u0131ndan do\u011frudan \u00f6\u011frenmek yerine, \u00f6nceden i\u015flenmi\u015f \u00f6zellikler veya farkl\u0131 karelerin birle\u015ftirilmesi gibi teknikler, a\u011f\u0131n \u00f6\u011frenmesi gereken karma\u015f\u0131kl\u0131\u011f\u0131 azaltarak sapmay\u0131 d\u00fc\u015f\u00fcrebilir ve varyans\u0131 kontrol edebilir. \u00d6rne\u011fin, son birka\u00e7 kareyi birle\u015ftirerek hareket bilgisini yakalamak.<\/p>\n<h5>2. D\u00fczenlile\u015ftirme (Regularization) Teknikleri (Varyans\u0131 Azaltma)<em><\/em><\/p>\n<p>*   <strong>A\u011f\u0131rl\u0131k Azaltma (Weight Decay &#8211; L1\/L2 Regularization):<\/strong> Sinir a\u011f\u0131n\u0131n a\u011f\u0131rl\u0131klar\u0131n\u0131n b\u00fcy\u00fckl\u00fc\u011f\u00fcn\u00fc cezaland\u0131rarak a\u015f\u0131r\u0131 uyumu \u00f6nler. Bu, daha basit bir modelin \u00f6\u011frenilmesine yard\u0131mc\u0131 olur.<br \/>\n*   <strong>Dropout:<\/strong> E\u011fitim s\u0131ras\u0131nda rastgele sinir h\u00fccresi n\u00f6ronlar\u0131n\u0131 devre d\u0131\u015f\u0131 b\u0131rak\u0131r. Bu, a\u011f\u0131n belirli n\u00f6ronlara a\u015f\u0131r\u0131 ba\u011f\u0131ml\u0131 olmas\u0131n\u0131 engelleyerek daha sa\u011flam \u00f6zelliklar \u00f6\u011frenmesini sa\u011flar.<br \/>\n*   <strong>Batch Normalization:<\/strong> Her katman\u0131n \u00e7\u0131kt\u0131s\u0131n\u0131 normalize ederek e\u011fitim s\u00fcrecini h\u0131zland\u0131r\u0131r ve kararl\u0131 hale getirir. Bu, daha y\u00fcksek \u00f6\u011frenme oranlar\u0131n\u0131n kullan\u0131lmas\u0131n\u0131 sa\u011flayarak sapmay\u0131 azaltmaya yard\u0131mc\u0131 olabilir ve ayn\u0131 zamanda d\u00fczenlile\u015ftirme etkisi de g\u00f6sterebilir.<\/p>\n<h5>3. Veri Art\u0131rma (Data Augmentation) (Varyans\u0131 Azaltma)<em><\/em><\/p>\n<p>*   Oyun ekranlar\u0131na rastgele \u00e7evirmeler, k\u0131rpmalar veya renk de\u011fi\u015fiklikleri uygulamak, ajan\u0131n farkl\u0131 varyasyonlara kar\u015f\u0131 daha dayan\u0131kl\u0131 olmas\u0131n\u0131 sa\u011flar. Bu, e\u011fitim verisinin \u00e7e\u015fitlili\u011fini art\u0131rarak modelin genelleme yetene\u011fini geli\u015ftirir ve varyans\u0131 azalt\u0131r.<\/p>\n<h5>4. E\u011fitim S\u00fcrecinin Y\u00f6netimi (Sapma ve Varyans \u00dczerindeki Etkileri)<em><\/em><\/p>\n<p>*   <strong>Erken Durdurma (Early Stopping):<\/strong> Do\u011frulama seti \u00fczerindeki performans\u0131 izleyerek, modelin performans\u0131 d\u00fc\u015fmeye ba\u015flad\u0131\u011f\u0131nda e\u011fitimi durdurmak, a\u015f\u0131r\u0131 uyumu \u00f6nler.<br \/>\n*   <strong>\u00d6\u011frenme Oran\u0131 Ayar\u0131 (Learning Rate Scheduling):<\/strong> \u00d6\u011frenme oran\u0131n\u0131 zamanla azaltmak, modelin daha ince ayarlamalar yapmas\u0131na ve yerel optimumlara daha iyi yak\u0131nsamas\u0131na yard\u0131mc\u0131 olabilir.<br \/>\n*   <strong>Toplu \u0130\u015f Boyutu (Batch Size):<\/strong> Daha b\u00fcy\u00fck toplu i\u015f boyutlar\u0131, gradyan tahminlerini daha kararl\u0131 hale getirerek varyans\u0131 azaltabilir, ancak daha k\u00fc\u00e7\u00fck toplu i\u015f boyutlar\u0131 daha iyi genelleme sa\u011flayabilir.<\/p>\n<h5>5. Algoritmik Yakla\u015f\u0131mlar (Sapma ve Varyans \u00dczerindeki Etkileri)<em><\/em><\/p>\n<p>*   <strong>Ke\u015fif Stratejileri (Exploration Strategies):<\/strong> Peki\u015ftirmeli \u00f6\u011frenmede, ajan sadece bildi\u011fi en iyi eylemleri se\u00e7mekle kalmamal\u0131, ayn\u0131 zamanda yeni eylemleri ke\u015ffetmelidir. epsilon-greedy gibi stratejiler, belirli bir olas\u0131l\u0131kla rastgele eylemler se\u00e7erek ke\u015ffi te\u015fvik eder. Yeterli ke\u015fif, ajan\u0131n daha iyi stratejiler bulmas\u0131n\u0131 sa\u011flayarak sapmay\u0131 azaltabilir. Ancak a\u015f\u0131r\u0131 ke\u015fif varyans\u0131 art\u0131rabilir.<br \/>\n*   <strong>Hedef A\u011flar (Target Networks) (DQN&#8217;de):<\/strong> DQN gibi algoritmalar, hedef de\u011ferleri hesaplamak i\u00e7in ayr\u0131 bir hedef a\u011f kullan\u0131r. Bu, hedef de\u011ferlerin daha kararl\u0131 olmas\u0131n\u0131 sa\u011flayarak e\u011fitim s\u00fcrecini stabilize eder ve varyans\u0131 azalt\u0131r.<br \/>\n*   <strong>Geri Yay\u0131l\u0131m (Backpropagation) ve Gradyan K\u0131rpma (Gradient Clipping):<\/strong> Sinir a\u011flar\u0131n\u0131n e\u011fitiminde kullan\u0131lan bu teknikler, b\u00fcy\u00fck gradyanlar\u0131n e\u011fitime zarar vermesini \u00f6nleyerek s\u00fcreci kararl\u0131 hale getirir ve varyans\u0131n kontrol alt\u0131nda tutulmas\u0131na yard\u0131mc\u0131 olur.<\/p>\n<p>### \u00d6rnek Uygulama: Basit Bir DQN Botu<\/p>\n<p>OpenAI Gym ile bir Atari botu olu\u015fturman\u0131n temel ad\u0131mlar\u0131n\u0131 ve sapma-varyans dengesini g\u00f6z \u00f6n\u00fcnde bulundurarak bir \u00f6rnek inceleyelim.<\/p>\n<p><strong>1. Kurulum:<\/strong><\/p>\n<pre><code class=\"language-python\">import gym\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout\nfrom tensorflow.keras.optimizers import Adam\nimport random\nfrom collections import deque<\/code><\/pre>\n<p><strong>2. Ortam\u0131n Haz\u0131rlanmas\u0131:<\/strong><\/p>\n<pre><code class=\"language-python\">env = gym.make(&quot;Pong-v0&quot;) # veya ba\u015fka bir Atari oyunu\nstate_shape = env.observation_space.shape\nnum_actions = env.action_space.n\n\n<h2>Durumlar\u0131 \u00f6n i\u015fleme fonksiyonu (\u00f6rne\u011fin, gri tonlama, yeniden boyutland\u0131rma, kareleri birle\u015ftirme)<\/h2>\ndef preprocess_state(state):\n    # Bu fonksiyon, oyunun durumunu sinir a\u011f\u0131n\u0131n anlayabilece\u011fi bir formata d\u00f6n\u00fc\u015ft\u00fcrmelidir.\n    # Genellikle, gri tonlama, yeniden boyutland\u0131rma ve son birka\u00e7 kareyi birle\u015ftirme yap\u0131l\u0131r.\n    # \u00d6rnek olarak, sadece yeniden boyutland\u0131rma:\n    state = tf.image.rgb_to_grayscale(state)\n    state = tf.image.resize(state, [84, 84])\n    return state.numpy().squeeze() # Squeeze ile tek boyutlu hale getir<\/code><\/pre>\n<p><strong>3. DQN Modeli Mimarisi (Sapma-Varyans Dengesi D\u00fc\u015f\u00fcn\u00fclerek):<\/strong><\/p>\n<pre><code class=\"language-python\">def build_dqn_model(input_shape, num_actions):\n    model = Sequential()\n    model.add(Conv2D(32, (8, 8), strides=(4, 4), activation='relu', input_shape=input_shape))\n    model.add(MaxPooling2D((2, 2), strides=(2, 2)))\n    model.add(Conv2D(64, (4, 4), activation='relu'))\n    model.add(MaxPooling2D((2, 2), strides=(2, 2)))\n    model.add(Conv2D(64, (3, 3), activation='relu'))\n    model.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\n    model.add(Flatten())\n    model.add(Dense(512, activation='relu'))\n    # Dropout ekleyerek varyans\u0131 azaltma\n    model.add(Dropout(0.3))\n    model.add(Dense(num_actions, activation='linear')) # Q-de\u011ferleri i\u00e7in linear aktivasyon\n\n    return model\n\n<h2>\u00d6rnek kullan\u0131m:<\/h2>\n<h2>input_shape = (84, 84, 4) # Son 4 kareyi birle\u015ftirdi\u011fimizi varsayal\u0131m<\/h2>\n<h2>model = build_dqn_model(input_shape, num_actions)<\/h2>\n<h2>model.compile(loss='mse', optimizer=Adam(learning_rate=0.0001))<\/code><\/pre>\n<\/h2>\n<p>*   Burada kullan\u0131lan CNN mimarisi, Atari oyunlar\u0131n\u0131n g\u00f6rsel girdilerini i\u015flemek i\u00e7in yayg\u0131n olarak kullan\u0131l\u0131r.<br \/>\n*   <code>Dropout<\/code> katman\u0131, e\u011fitim s\u0131ras\u0131nda rastgele n\u00f6ronlar\u0131 devre d\u0131\u015f\u0131 b\u0131rakarak a\u015f\u0131r\u0131 uyumu engeller, yani varyans\u0131 azalt\u0131r.<br \/>\n*   D\u00fc\u015f\u00fck bir \u00f6\u011frenme oran\u0131 (<code>0.0001<\/code>) ile Adam optimizasyonunu kullanmak, daha kararl\u0131 bir \u00f6\u011frenme s\u00fcreci sa\u011flar ve sapma\/varyans sorunlar\u0131n\u0131 y\u00f6netmeye yard\u0131mc\u0131 olur.<\/p>\n<p><strong>4. DQN Algoritmas\u0131 (Temel Yap\u0131):<\/strong><\/p>\n<pre><code class=\"language-python\">class DQNAgent:\n    def __init__(self, state_shape, num_actions, gamma=0.99, epsilon=1.0, epsilon_decay=0.995, epsilon_min=0.01, replay_buffer_size=10000, batch_size=32, target_update_freq=10):\n        self.state_shape = state_shape\n        self.num_actions = num_actions\n        self.gamma = gamma\n        self.epsilon = epsilon\n        self.epsilon_decay = epsilon_decay\n        self.epsilon_min = epsilon_min\n        self.batch_size = batch_size\n        self.target_update_freq = target_update_freq\n        self.memory = deque(maxlen=replay_buffer_size)\n        self.learning_rate = 0.0001\n\n        self.model = build_dqn_model(state_shape, num_actions)\n        self.target_model = build_dqn_model(state_shape, num_actions)\n        self.target_model.set_weights(self.model.get_weights()) # Ba\u015flang\u0131\u00e7ta a\u011f\u0131rl\u0131klar\u0131 senkronize et\n\n        self.optimizer = Adam(learning_rate=self.learning_rate)\n        self.model.compile(loss='mse', optimizer=self.optimizer)\n\n    def remember(self, state, action, reward, next_state, done):\n        self.memory.append((state, action, reward, next_state, done))\n\n    def choose_action(self, state):\n        if np.random.rand() &lt;= self.epsilon:\n            return random.randrange(self.num_actions)\n        else:\n            state = np.expand_dims(state, axis=0)\n            q_values = self.model.predict(state)\n            return np.argmax(q_values[0])\n\n    def replay(self, episode_count):\n        if len(self.memory) &lt; self.batch_size:\n            return\n\n        minibatch = random.sample(self.memory, self.batch_size)\n\n        states = np.array([i[0] for i in minibatch])\n        actions = np.array([i[1] for i in minibatch])\n        rewards = np.array([i[2] for i in minibatch])\n        next_states = np.array([i[3] for i in minibatch])\n        dones = np.array([i[4] for i in minibatch])\n\n        # Hedef Q de\u011ferlerini hesapla\n        target_q_next = self.target_model.predict(next_states)\n        target_q = rewards + self.gamma <em> np.max(target_q_next, axis=1) <\/em> (1 - dones)\n\n        # Mevcut Q de\u011ferlerini hesapla\n        current_q = self.model.predict(states)\n\n        # Hedef Q de\u011ferlerini g\u00fcncelle\n        for i in range(self.batch_size):\n            current_q[i][actions[i]] = target_q[i]\n\n        # Modeli e\u011fit\n        self.model.fit(states, current_q, epochs=1, verbose=0)\n\n        # Epsilon'u azalt\n        if self.epsilon &gt; self.epsilon_min:\n            self.epsilon *= self.epsilon_decay\n\n        # Hedef modeli g\u00fcncelleme\n        if episode_count % self.target_update_freq == 0:\n            self.target_model.set_weights(self.model.get_weights())<\/code><\/pre>\n<p>*   <strong>Replay Buffer (Bellek):<\/strong> Deneyimleri rastgele \u00f6rnekleyerek veri korelasyonunu azalt\u0131r ve varyans\u0131 d\u00fc\u015f\u00fcrmeye yard\u0131mc\u0131 olur.<br \/>\n*   <strong>Target Network:<\/strong> Hedef de\u011ferlerin hesaplanmas\u0131n\u0131 stabilize ederek sapma ve varyans\u0131n y\u00f6netimini kolayla\u015ft\u0131r\u0131r.<br \/>\n*   <strong>Epsilon Decay:<\/strong> Ba\u015flang\u0131\u00e7ta daha fazla ke\u015fif (y\u00fcksek epsilon) yaparak sapmay\u0131 azaltmay\u0131 hedefler, sonra ke\u015ffi azalt\u0131p \u00f6\u011frenilen bilgilere odaklanarak (d\u00fc\u015f\u00fck epsilon) varyans\u0131 kontrol alt\u0131nda tutar.<\/p>\n<p><strong>5. E\u011fitim D\u00f6ng\u00fcs\u00fc:<\/strong><\/p>\n<pre><code class=\"language-python\"># E\u011fitim parametreleri\nnum_episodes = 1000\n<h2>... (Di\u011fer DQNAgent parametreleri)<\/h2>\n\nagent = DQNAgent(state_shape=(84, 84, 4), num_actions=num_actions) # State shape'i uygun \u015fekilde ayarlay\u0131n\n\nfor episode in range(num_episodes):\n    state = env.reset()\n    # \u0130lk durumu \u00f6n i\u015fleme ve yeterli kare ekleme\n    processed_state = preprocess_state(state)\n    state_buffer = deque([processed_state] * 4, maxlen=4) # Son 4 kareyi sakla\n    current_state = np.stack(state_buffer, axis=-1)\n\n    total_reward = 0\n    done = False\n    step = 0\n\n    while not done:\n        action = agent.choose_action(current_state)\n        next_state, reward, done, _ = env.step(action)\n        processed_next_state = preprocess_state(next_state)\n        state_buffer.append(processed_next_state)\n        next_state_stacked = np.stack(state_buffer, axis=-1)\n\n        agent.remember(current_state, action, reward, next_state_stacked, done)\n        agent.replay(episode)\n\n        current_state = next_state_stacked\n        total_reward += reward\n        step += 1\n\n        if done:\n            print(f&quot;Episode: {episode}, Total Reward: {total_reward}, Epsilon: {agent.epsilon:.2f}&quot;)\n            break<\/code><\/pre>\n<p>Bu \u00f6rnek, temel bir DQN yap\u0131s\u0131n\u0131 g\u00f6stermektedir. Ger\u00e7ek bir Atari botu geli\u015ftirmek i\u00e7in durum \u00f6n i\u015fleme, bellek y\u00f6netimi, hedef a\u011f g\u00fcncellemeleri ve hiperparametre ayarlar\u0131 gibi bir\u00e7ok detay\u0131n dikkatlice yap\u0131lmas\u0131 gerekir.<\/p>\n<h3>Sonu\u00e7<\/h3>\n<p>Derin peki\u015ftirmeli \u00f6\u011frenmede sapma-varyans dengesi, modelin genelleme yetene\u011fini ve performans\u0131n\u0131 do\u011frudan etkileyen kritik bir fakt\u00f6rd\u00fcr. OpenAI Gym gibi ara\u00e7larla Atari botlar\u0131 in\u015fa ederken, bu dengeyi anlamak ve y\u00f6netmek, ba\u015far\u0131l\u0131 bir ajan geli\u015ftirmek i\u00e7in \u015fartt\u0131r. Sinir a\u011f\u0131 mimarisini dikkatlice se\u00e7mek, d\u00fczenlile\u015ftirme tekniklerini uygulamak, veri art\u0131rma y\u00f6ntemlerinden yararlanmak ve e\u011fitim s\u00fcrecini optimize etmek gibi stratejiler, hem sapmay\u0131 d\u00fc\u015f\u00fck tutmaya hem de varyans\u0131 kontrol alt\u0131nda tutmaya yard\u0131mc\u0131 olur. Bu makalede sunulan kavramlar ve \u00f6rnek, DRL projelerinde sapma-varyans dengesini daha bilin\u00e7li bir \u015fekilde ele almak i\u00e7in bir temel olu\u015fturmaktad\u0131r. Bu prensipleri uygulayarak, daha sa\u011flam, genellenebilir ve y\u00fcksek performansl\u0131 DRL ajanlar\u0131 in\u015fa edebiliriz.<\/p>\n<div class=\"github-example-link\"><strong>\u00d6rnek kod:<\/strong> <a href=\"https:\/\/github.com\/fatihsoysalcom\/bias-variance-tradeoff-atari-bot\" target=\"_blank\" rel=\"noopener noreferrer\">github.com\/fatihsoysalcom\/bias-variance-tradeoff-atari-bot<\/a><\/div>\n","protected":false},"excerpt":{"rendered":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme Derin Peki\u015ftirmeli \u00d6\u011frenme (Deep Reinforcement Learning &#8211; DRL), bilgisayar sistemlerinin karma\u015f\u0131k ortamlarda insan benzeri veya \u00fcst\u00fcn performans sergilemesini sa\u011flayan g\u00fc\u00e7l\u00fc bir yapay zeka alan\u0131d\u0131r.","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1342],"tags":[],"class_list":{"0":"post-42817","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-ai","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v20.5 (Yoast SEO v25.3.1) - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme - Kodlar\u0131n Gizemli D\u00fcnyas\u0131<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\" \/>\n<meta property=\"og:locale\" content=\"tr_TR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme\" \/>\n<meta property=\"og:description\" content=\"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme Derin Peki\u015ftirmeli \u00d6\u011frenme (Deep Reinforcement Learning - DRL), bilgisayar sistemlerinin karma\u015f\u0131k ortamlarda insan benzeri veya \u00fcst\u00fcn performans sergilemesini sa\u011flayan g\u00fc\u00e7l\u00fc bir yapay zeka alan\u0131d\u0131r.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\" \/>\n<meta property=\"og:site_name\" content=\"Kodlar\u0131n Gizemli D\u00fcnyas\u0131\" \/>\n<meta property=\"article:published_time\" content=\"2026-06-24T14:02:55+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-06-24T14:04:24+00:00\" \/>\n<meta name=\"author\" content=\"Fatih Soysal\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Yazan:\" \/>\n\t<meta name=\"twitter:data1\" content=\"Fatih Soysal\" \/>\n\t<meta name=\"twitter:label2\" content=\"Tahmini okuma s\u00fcresi\" \/>\n\t<meta name=\"twitter:data2\" content=\"14 dakika\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\"},\"author\":{\"name\":\"Fatih Soysal\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"headline\":\"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme\",\"datePublished\":\"2026-06-24T14:02:55+00:00\",\"dateModified\":\"2026-06-24T14:04:24+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\"},\"wordCount\":2020,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"articleSection\":[\"AI\"],\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#respond\"]}],\"copyrightYear\":\"2026\",\"copyrightHolder\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#organization\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\",\"name\":\"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme - Kodlar\u0131n Gizemli D\u00fcnyas\u0131\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\"},\"datePublished\":\"2026-06-24T14:02:55+00:00\",\"dateModified\":\"2026-06-24T14:04:24+00:00\",\"breadcrumb\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#breadcrumb\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Anasayfa\",\"item\":\"https:\/\/fatihsoysal.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/\",\"name\":\"Fatihsoysal.com\",\"description\":\"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim\",\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"tr\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\",\"name\":\"Fatih Soysal\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"tr\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"contentUrl\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"width\":512,\"height\":512,\"caption\":\"Fatih Soysal\"},\"logo\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\"},\"description\":\"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/","og_locale":"tr_TR","og_type":"article","og_title":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme","og_description":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme Derin Peki\u015ftirmeli \u00d6\u011frenme (Deep Reinforcement Learning - DRL), bilgisayar sistemlerinin karma\u015f\u0131k ortamlarda insan benzeri veya \u00fcst\u00fcn performans sergilemesini sa\u011flayan g\u00fc\u00e7l\u00fc bir yapay zeka alan\u0131d\u0131r.","og_url":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/","og_site_name":"Kodlar\u0131n Gizemli D\u00fcnyas\u0131","article_published_time":"2026-06-24T14:02:55+00:00","article_modified_time":"2026-06-24T14:04:24+00:00","author":"Fatih Soysal","twitter_card":"summary_large_image","twitter_misc":{"Yazan:":"Fatih Soysal","Tahmini okuma s\u00fcresi":"14 dakika"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#article","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/"},"author":{"name":"Fatih Soysal","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"headline":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme","datePublished":"2026-06-24T14:02:55+00:00","dateModified":"2026-06-24T14:04:24+00:00","mainEntityOfPage":{"@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/"},"wordCount":2020,"commentCount":0,"publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"articleSection":["AI"],"inLanguage":"tr","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#respond"]}],"copyrightYear":"2026","copyrightHolder":{"@id":"https:\/\/fatihsoysal.com\/blog\/#organization"}},{"@type":"WebPage","@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/","url":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/","name":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/#website"},"datePublished":"2026-06-24T14:02:55+00:00","dateModified":"2026-06-24T14:04:24+00:00","breadcrumb":{"@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#breadcrumb"},"inLanguage":"tr","potentialAction":[{"@type":"ReadAction","target":["https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/fatihsoysal.com\/blog\/derin-pekistirmeli-ogrenme-icin-sapma-varyans-dengesi-openai-gym-ile-atari-botu-insa-etme\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Anasayfa","item":"https:\/\/fatihsoysal.com\/blog\/"},{"@type":"ListItem","position":2,"name":"Derin Peki\u015ftirmeli \u00d6\u011frenme i\u00e7in Sapma-Varyans Dengesi: OpenAI Gym ile Atari Botu \u0130n\u015fa Etme"}]},{"@type":"WebSite","@id":"https:\/\/fatihsoysal.com\/blog\/#website","url":"https:\/\/fatihsoysal.com\/blog\/","name":"Fatihsoysal.com","description":"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim","publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"tr"},{"@type":["Person","Organization"],"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1","name":"Fatih Soysal","image":{"@type":"ImageObject","inLanguage":"tr","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/","url":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","contentUrl":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","width":512,"height":512,"caption":"Fatih Soysal"},"logo":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/"},"description":"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.","url":"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/"}]}},"yoast_meta":{"yoast_wpseo_title":"","yoast_wpseo_metadesc":"","yoast_wpseo_canonical":""},"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/42817","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/comments?post=42817"}],"version-history":[{"count":2,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/42817\/revisions"}],"predecessor-version":[{"id":42819,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/42817\/revisions\/42819"}],"wp:attachment":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/media?parent=42817"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/categories?post=42817"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/tags?post=42817"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}