Jiu chess, as an intangible cultural heritage of the Tibetan ethnic group, possesses unique rules and cultural significance. However, research on its computer game faces challenges such as scarce game data and complex rules. This study proposes Jiu-E, an expert knowledge-guided deep reinforcement learning model, to enhance the game-playing capabilities of Jiu chess agents. First, based on regional control theory and offensive-defensive evaluation mechanisms, an expert-knowledge-driven opening book is designed to ensure strategic advantages during the battle stage. Second, an action evaluation framework is constructed, incorporating stage-specific action-value functions for the layout and battle stages, along with an innovative sliding window-based endgame evaluation algorithm to address the sparsity and dynamic challenges in endgame scenarios. Finally, a supervised pre-training model is developed and enhanced by integrating expert knowledge-guided node reward functions to mitigate the reward sparsity in deep reinforcement learning for Jiu chess. Experimental results demonstrate that Jiu-E achieves a significantly higher win rate compared to other models. Ablation studies further validate the effectiveness of the node reward function. This work provides novel insights for intelligent game research on culturally significant chess variants and advances the integration of artificial intelligence with traditional heritage.