We develop an approximate dynamic programming approach to network revenue management models with customer choice that approximates the value function of the Markov decision process with a non-linear function which is separable across resource inventory levels. This approximation can exhibit significantly improved accuracy compared to currently available methods. It further allows for arbitrary aggregation of inventory units and thereby reduction of computational workload, yields upper bounds on the optimal expected revenue that are provably at least as tight as those obtained from previous approaches, and is asymptotically optimal under fluid scaling. Computational experiments for the multinomial logit choice model with distinct consideration sets show that policies derived from our approach outperform available alternatives, and we demonstrate how aggregation can be used to balance solution quality and runtime.