In this paper we develop a link cost function for data delivery in sensor network. Usually most conventional methods determine the optimal coefficients in the cost function without considering the surrounding environment of the node such as the wireless propagation environment or the topological environment. Due to this reason, there are limitations to improve the quality of data delivery such as data delivery ratio and delay of data delivery. To solve this problem, we derive a new cost function using the concept of Partially Connected Neural Network (PCNN) which is modeled according to the input types whether inputs are correlated or uncorrelated. The correlated inputs are connected to the hidden layer of the PCNN in a coupled fashion but the uncoupled inputs are in an uncoupled fashion. We also propose the training technique for finding an optimal weight vector in the link cost function. The link cost function is trained to the direction that the packet transmission success ratio of each node maximizes. In the experimental section, we show that our method outperforms other conventional methods in terms of the quality of data delivery and the energy efficiency.