This paper introduces MetaQNN, a reinforcement learning-based algorithm for automatically generating high-performing convolutional neural network (CNN) architectures. The algorithm uses Q-learning with an ε-greedy exploration strategy and experience replay to sequentially select CNN layers. The agent explores a large but finite space of possible architectures and iteratively discovers designs with improved performance on the learning task. On image classification benchmarks, the agent-designed networks (consisting of only standard convolution, pooling, and fully-connected layers) beat existing networks designed with the same layer types and are competitive against state-of-the-art methods that use more complex layer types. The MetaQNN method outperforms existing meta-modeling approaches for network design on image classification tasks. The paper discusses the design of neural network architectures, including the use of reinforcement learning for automated design. It also covers related work in Bayesian optimization and reinforcement learning for deep learning. The paper presents a detailed background on Q-learning, the theoretical formulation of Q-learning, and the recursive maximization equation known as Bellman's Equation. The paper also describes the design choices and learning process in detail, including the state space, action space, and Q-learning training procedure. The paper presents experiment details, including the training process, validation set creation, and hyperparameter tuning. The results show that the MetaQNN networks outperform previous meta-modeling methods and hand-crafted networks which use the same types of layers. The paper also discusses the transfer learning ability of the MetaQNN networks and concludes that the method is able to generate tailored CNN designs for different image classification tasks. The paper also discusses future research directions in reinforcement learning-driven network design.This paper introduces MetaQNN, a reinforcement learning-based algorithm for automatically generating high-performing convolutional neural network (CNN) architectures. The algorithm uses Q-learning with an ε-greedy exploration strategy and experience replay to sequentially select CNN layers. The agent explores a large but finite space of possible architectures and iteratively discovers designs with improved performance on the learning task. On image classification benchmarks, the agent-designed networks (consisting of only standard convolution, pooling, and fully-connected layers) beat existing networks designed with the same layer types and are competitive against state-of-the-art methods that use more complex layer types. The MetaQNN method outperforms existing meta-modeling approaches for network design on image classification tasks. The paper discusses the design of neural network architectures, including the use of reinforcement learning for automated design. It also covers related work in Bayesian optimization and reinforcement learning for deep learning. The paper presents a detailed background on Q-learning, the theoretical formulation of Q-learning, and the recursive maximization equation known as Bellman's Equation. The paper also describes the design choices and learning process in detail, including the state space, action space, and Q-learning training procedure. The paper presents experiment details, including the training process, validation set creation, and hyperparameter tuning. The results show that the MetaQNN networks outperform previous meta-modeling methods and hand-crafted networks which use the same types of layers. The paper also discusses the transfer learning ability of the MetaQNN networks and concludes that the method is able to generate tailored CNN designs for different image classification tasks. The paper also discusses future research directions in reinforcement learning-driven network design.