This paper presents CoDeepNEAT, an automated method for evolving deep neural network (DNN) architectures. The method extends neuroevolution techniques to optimize network topology, components, and hyperparameters, achieving results comparable to human-designed architectures in standard benchmarks for object recognition and language modeling. It also supports building a real-world application for automated image captioning on a magazine website. The approach is computationally intensive but becomes more effective with increased computing power, making it a promising method for future deep learning applications. The paper discusses the challenges of designing complex DNNs, where human engineers can only optimize a few parameters, while DNNs have complex topologies and many hyperparameters. The complexity challenge is not unique to neural networks, and new approaches are emerging where humans handle high-level design, while computational systems optimize details. This approach is applied to DNN architecture design, which includes three challenges: designing components, assembling them into a network topology, and setting hyperparameters. The paper develops an approach for automatic DNN design based on the neuroevolution technique NEAT, which has been successful in evolving small recurrent networks. NEAT is extended to coevolution of components, topologies, and hyperparameters. The fitness of evolved networks is determined based on their ability to perform well in the task through gradient descent. The approach is demonstrated in standard benchmark tasks of object recognition and language modeling, and in a real-world application of captioning images on a magazine website. The results show that the approach discovers designs comparable to the state of the art, and does it automatically without much development effort. The approach is computationally extremely demanding, but with more computational power, it is likely to be more effective and possibly surpass human design. Such power is now becoming available in various forms of cloud computing and grid computing, making evolutionary optimization of neural networks a promising approach for the future. The paper also discusses the evolution of LSTM architectures, extending CoDeepNEAT to search for connectivity between LSTM layers. The approach is evaluated in the standard benchmark task of language modeling, showing that even small innovations in components can have a significant effect on performance. The paper concludes that evolutionary optimization makes it possible to construct more complex deep learning architectures than can be done by hand. The topology, components, and hyperparameters of the architecture can all be optimized simultaneously to fit the requirements of the task, resulting in superior performance. This automated design can make new applications of deep learning possible in vision, speech, language, and other areas. Currently such designs are comparable with best human designs; with anticipated increases in computing power, they should soon surpass them, putting the power to good use.This paper presents CoDeepNEAT, an automated method for evolving deep neural network (DNN) architectures. The method extends neuroevolution techniques to optimize network topology, components, and hyperparameters, achieving results comparable to human-designed architectures in standard benchmarks for object recognition and language modeling. It also supports building a real-world application for automated image captioning on a magazine website. The approach is computationally intensive but becomes more effective with increased computing power, making it a promising method for future deep learning applications. The paper discusses the challenges of designing complex DNNs, where human engineers can only optimize a few parameters, while DNNs have complex topologies and many hyperparameters. The complexity challenge is not unique to neural networks, and new approaches are emerging where humans handle high-level design, while computational systems optimize details. This approach is applied to DNN architecture design, which includes three challenges: designing components, assembling them into a network topology, and setting hyperparameters. The paper develops an approach for automatic DNN design based on the neuroevolution technique NEAT, which has been successful in evolving small recurrent networks. NEAT is extended to coevolution of components, topologies, and hyperparameters. The fitness of evolved networks is determined based on their ability to perform well in the task through gradient descent. The approach is demonstrated in standard benchmark tasks of object recognition and language modeling, and in a real-world application of captioning images on a magazine website. The results show that the approach discovers designs comparable to the state of the art, and does it automatically without much development effort. The approach is computationally extremely demanding, but with more computational power, it is likely to be more effective and possibly surpass human design. Such power is now becoming available in various forms of cloud computing and grid computing, making evolutionary optimization of neural networks a promising approach for the future. The paper also discusses the evolution of LSTM architectures, extending CoDeepNEAT to search for connectivity between LSTM layers. The approach is evaluated in the standard benchmark task of language modeling, showing that even small innovations in components can have a significant effect on performance. The paper concludes that evolutionary optimization makes it possible to construct more complex deep learning architectures than can be done by hand. The topology, components, and hyperparameters of the architecture can all be optimized simultaneously to fit the requirements of the task, resulting in superior performance. This automated design can make new applications of deep learning possible in vision, speech, language, and other areas. Currently such designs are comparable with best human designs; with anticipated increases in computing power, they should soon surpass them, putting the power to good use.