The study proposes a novel framework for portfolio optimization by combining the Black-Litterman model and deep reinforcement learning. The Black-Litterman model provides posterior returns based on market equilibrium and investor views, thus providing a stable platform for onward decision-making. Taking these outputs and using them in a Markov Decision Process framework, the system will dynamically update the portfolio allocation to ensure risk-return trade-offs in volatile markets. The model proposed here integrates an improved reward function that includes a Sharpe ratio in assessing risk-adjusted performance, thus leading the agent to optimize long/short strategies. Back-testing on major U.S. technology stocks shows that this framework performs better than equal-weighted portfolios and mean-variance portfolios in terms of cumulative return, Sharpe ratio, and drawdown. This result points to the potential of combining machine learning with financial theories to address the complexities of portfolio management and the reality of change in market conditions.