AI software is regarded as the key to the formation of new domain and new quality combat capabilities in the future. In response to the characteristics of randomness, autonomy, and learning of AI software, traditional testing methods lack comprehensive coverage of non-functional requirements, test data is difficult to obtain and annotate, and evaluation indicators are not comprehensive. This paper focuses on the research of intelligent algorithm evaluation technology, proposing a multi-perspective and multi-attribute evaluation method for intelligent algorithms from four aspects: accuracy evaluation, efficiency evaluation, data robustness evaluation and algorithm coverage evaluation. This can effectively support future AI software evaluation work and improve the AI software testing technology system.