Historically, deep learning has proven increasingly useful and has become a frequently used tool for a wide range of applications. However, the sensitivity of deep learning systems to the specific network architecture, both in terms of learning ability and performance, has made them difficult to utilize for non-experts and emphasized the need for automatic network architecture optimization. Previous work largely optimizes a network for one specific problem using architecture search, requiring significant amounts of time training different architectures during optimization. To alleviate this, the approach proposed here learns a network optimization policy on an abstract problem and architecture embedding those transfers across problems and thus permits the optimization of networks for new problems without excessive additional training. Siamese networks here learn an effective, coherent embedding of the network architecture space, which, together with a problem-specific feature vector to capture the complexity of the problem, serve as a basis for a reinforcement learning agent to acquire a transferable policy to derive high-performing network architectures for a large array of problems. To demonstrate the feasibility and potential of the approach, initial experiments with a set of standard classification problems and an architecture space comprising a limited range of fully connected networks are performed, showing the ability of the approach to learn a policy that optimizes the architecture even for new problems that were not part of policy training. In summary, we employed a novel RL approach to successfully learn a policy for network optimization in a derived architecture embedding space that incrementally optimized the network.