In recent years, a new learn-to-invest framework using direct investment performance optimization techniques has emerged and is gradually gaining recognition as a promising framework for develiping intelligent investment systems. This methodology continues earlier efforts in which similar investment problems are formulated from the standpoint of traditional dynamic programming and stochastic control. In this paper, we propose to train an S&P 500/ T-bill asset allocation system by optimizing the utility function directly through reinforcement learning techniques. The preseuted novel approach is theoretically appealin due to the fact that it is a one-step optimization process and it does not require any intermediate steps, such as making forecasts or labeling desired investments. The simulation results demonstrate the effectiveness of this asset allocation strategy.