With the rapid development of Internet technology, recommendation system has become an important tool to improve user experience and increase platform revenue. In recent years, Deep Reinforcement Learning (DRL) has become an important method in the field of recommendation system optimization with its dynamic learning and policy optimization capabilities. However, the optimization goals of the recommendation system are diverse, including click rate, user retention rate, diversity of recommended content, etc., and a single optimization goal is difficult to fully meet the needs of the platform. In this paper, an optimization method of deep reinforcement learning recommendation system based on comprehensive reward function is proposed, which balances multiple optimization objectives by designing multi-objective comprehensive reward function. By constructing a comprehensive reward function that includes click-through rate, user retention rate and diversity of recommended content, Deep reinforcement learning algorithms such as deep Q-Network (DQN) are used to continuously optimize recommendation strategies. The comprehensive reward function combines different objectives by weighted summation, which makes the system improve the recommendation effect in multiple dimensions. The experiment uses real user data of an e-commerce platform for verification. Compared with other methods such as collaborative filtering, the recommendation accuracy of this method is improved by more than 5%, the content diversity is improved, and the user retention rate is increased by about 3%. By balancing multiple optimization objectives, deep reinforcement learning achieves superior results in improving the user experience and overall system performance.