Training of combat fighter pilots is often conducted using either human opponents or non-adaptive computer-generated force (CGF) inserted with the doctrine for conducting air combat mission. The novelty and challenges of such non-adaptive doctrine-driven CGF is often lost quickly. Incorporating more complex knowledge manually is known to be tedious and time-consuming. Therefore, a study of using adaptive CGF to learn from the real-time interactions with human pilots to extend the existing doctrine is conducted in this work. The goal of this study is to show how an adaptive CGF can be more effective than a non-adaptive doctrine-driven CGF for simulator-based training of combat pilots. Driven by a family of self-organizing neural network, the adaptive CGF can be inserted with the same doctrine as the non-adaptive CGF. Using a commercial-grade training simulation platform, two human-in-the-loop (HIL) experiments are conducted using the adaptive CGF and the non-adaptive doctrine-driven CGF to engage two diverse groups of human pilots in 1-v-1 dogfights. The quantitative results and qualitative assessments of the CGFs by the human pilots are collected for all the training sessions. The qualitative assessments show the trainee pilots are able to match the adaptive CGF to the desirable attributes while the veteran pilots are only able to observe some learning from the adaptive CGF. The quantitative results show that the adaptive agent needs a lot more training sessions to learn the necessary knowledge to match up to the human pilots.