{"id":909,"date":"2018-10-01T03:17:08","date_gmt":"2018-10-01T03:17:08","guid":{"rendered":"http:\/\/kusuaks7\/?p=514"},"modified":"2023-07-25T16:45:49","modified_gmt":"2023-07-25T16:45:49","slug":"the-secrets-behind-reinforcement-learning","status":"publish","type":"post","link":"https:\/\/www.experfy.com\/blog\/ai-ml\/the-secrets-behind-reinforcement-learning\/","title":{"rendered":"The secrets behind Reinforcement Learning"},"content":{"rendered":"

Ready to learn Machine Learning? Browse<\/em><\/strong> Machine Learning Training and Certification courses<\/a> developed by industry thought leaders and Experfy in Harvard Innovation Lab.<\/em><\/strong><\/p>\n

Bots that play Dota2, AI that beat the best Go players in the world, computers that excel at Doom. What\u2019s going on? Is there a reason why the AI community has been so busy playing games?<\/p>\n

Let me put it that way. If you want a robot to learn how to walk what do you do? You build one, program it and release it on the streets of New York? Of course not. You build a simulation, a game, and you use that virtual space to teach it how to move around it. Zero cost, zero risks. That\u2019s why games are so useful in research areas. But how do you teach it to walk? The answer is the topic of today\u2019s article and is probably the most exciting field of Machine learning at the time:<\/p>\n

You probably knew that there are two types of machine learning. Supervised and unsupervised. Well, there is a third one, called Reinforcement Learning. RL is arguably the most difficult area of ML to understand cause there are so many, many things going on at the same time. I\u2019ll try to simplify as much as I can because it is a really astonishing area and you should definitely know about it. But let me warn you. It involves complex thinking and 100% focus to grasp it. And some math. So, take a deep breath and let\u2019s dive in:<\/p>\n

Markov Decision Processes<\/h3>\n

Reinforcement learning is a trial and error process where an AI (agent<\/strong>) performs a number of\u00a0actions<\/strong>\u00a0in an\u00a0environment<\/strong>. Each unique moment the agent has a\u00a0state<\/strong>\u00a0and acts from this given state to a new one. This particular action may on may not has a\u00a0reward.<\/strong>\u00a0Therefore, we can say that each learning epoch (or episode) can be represented as a sequence of states, actions, and rewards. Each state depends only on the previous states and actions and as the environment is inherently stochastic (we don\u2019t know the state that comes next), this process satisfies the\u00a0Markov property<\/a>. The Markov property says that the conditional probability distribution of future states of the process depends only upon the present state, not on the sequence of events that preceded it. The whole process is referred to as\u00a0a Markov Decision Process<\/a>. Markov Decision Process is the main mathematical tool we use to frame almost any RL problem in a way that is easy to study and experimenting different solutions.<\/p>\n

<\/canvas><\/figure>\n

Let\u2019s see a real example using Super Mario. In this case:<\/p>\n