In multi-state environments, Q-learning reaches its limits. Deep Q-Learning uses neural networks to efficiently solve a problem.
reading time:
23 Min.
In Pocket save on computer
Created with Midjourney by heise online.
An artificial intelligence (AI) like ChatGPT uses neural networks to independently find the most suitable answer to an input. But how exactly does a neural network work? And how can a user use it to develop their own self-learning system? In this article, we explain in detail the concepts of Deep Q Learning, which uses a neural network.
We show how deep-Q-learning works using the CartPole problem: The AI should balance a virtual pole on a cart and carry out the best possible action in every situation so that the pole does not fall over. We explain different aspects of neural networks and how to use them in deep-Q-learning. We also describe fixed Q targets and replay buffers.
This is what the CartPole problem looks like: A pole balances on a cart that the AI can control. Now the stick must not fall over.
Deep Q-Learning is not only interesting for AI developers, but also for anyone who wants to understand how machines can develop self-learning systems through trial and error that are able to handle complex tasks. Since we are going deep into the matter with this article, we assume basic knowledge of Python and machine learning. If you still need a quick refresher, you will find easy access to the topic and explanations for many of the technical terms used here in our introductory article on Q-Learning.
