Portfolio offers an effective way for managing investment risk through diversification. The key issue in portfolio management is how to determine the weight (portion) of each asset in the portfolio, so as to achieve high profit with low risk over a certain period of trading. We propose a learning-based trading strategy for portfolio management, which aims at maximizing the Sharpe Ratio by actively reallocating wealth among assets. The trading decision is formulated as a non-linear function of the latest realized asset returns, and the function can be approximated by a neural-network. Two methods based on supervised learning to train the network are proposed. Experiments show that the proposed trading strategy outperforms the static Sharpe Ratio trading method.