Portfolio allocation aims to optimize the risk/return ratio in investment management. Traditional methods based on modern portfolio theory have been widely used for this purpose. However, the emergence of deep reinforcement learning (DRL) offers an alternative approach. This article conducts a comprehensive comparative analysis of traditional portfolio allocation methods and DRL, examining their principles, methodologies, and performance in maximizing risk-return profiles. It demonstrates that a basic version of DRL converges to traditional methods, while a myopic agent driven by immediate rewards represents the dynamic version of traditional methods. Experimental results indicate some improvement of DRL over traditional methods.