Portfolio optimization faces challenges from non-stationary market dynamics and the limitations of existing methods in handling complex data relationships and preserving prior knowledge. Standard Deep Reinforcement Learning (DRL) approaches often rely on static rewards, struggle to integrate multiple view types beyond expected returns, and discard historical information non-optimally. This paper introduces the Deep Multi-View Factor Entropy Pooling (DMVFEP) framework to address these issues. DMVFEP utilizes a novel multi-view architecture with specialized neural networks including Transformer for returns, asymmetric Multilayer Perceptron (MLP) with attention for volatility, and MLP for factors, integrated into a Twin Delayed Deep Deterministic Policy Gradient DRL agent. Crucially, the agent learns to generate dynamic, state-dependent market views, rather than portfolio weights directly. These views are then systematically integrated with prior beliefs using Factor Entropy Pooling, a principled method that minimizes information loss. Empirical evaluation on Dow Jones Industrial Average-derived assets demonstrates DMVFEP's outstanding outperformance, yielding significantly higher absolute and risk-adjusted returns compared to benchmark strategies and leading DRL models, achieving an improvement of 56% compared to the next best performing DRL model. Ablation studies confirm the significant contribution of each view component.