Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder that affects children's social interactions, communication skills, and behavior patterns. Music therapy, as a non-verbal communication method, has been widely recognized to have a positive therapeutic effect on children with autism. However, traditional music therapy lacks personalization and adaptability, and it is difficult to meet the unique needs of each child. This paper aims to develop an intelligent adaptive music therapy system for children with autism based on reinforcement learning. First, the basic principles of reinforcement learning technology and its application potential in the medical field are introduced. Next, the theoretical basis and current practice of music therapy for autistic children are elaborated in detail. Then, the overall architecture of the system is proposed, including modules such as music selection, rhythm adjustment, and interactive feedback, and how to use reinforcement learning algorithms to achieve intelligent adaptability of the system is explained in detail. Finally, the effectiveness and adaptability of the system are verified through practical experiments. Experimental results show that the system can intelligently adjust the music therapy plan according to the real-time behavioral responses of autistic children, improve their psychological state, and alleviate the social disorders of autistic children. The degree of their social disorders is significantly reduced and shows a stable downward trend, with the lowest value reaching 0.2.