This work is centered around advanced Imitation Learning (IL) techniques to develop an agent capable of imitating human behavior in a videogame, where behavior changes with the mastery of the game. A common problem in supervised approaches is the acquisition of labeled data, which is often costly and can also be subject to concept drift. This work explores improvements in the adaptability and effectiveness of learning an expert policy in dynamic and complex environments such as video games using Behavioral Cloning (BC) and Meta-learning. We propose a learner based on a hybrid approach, utilizing both BC and a blending ensemble of deep models to enhance the agent’s generalization capability on unseen trajectories with minimal human intervention. Results indicate that this approach improves training effectiveness and achieves comparable classification performance to a traditional DAgger approach with less data training, improving the base learner’s training effectiveness and adaptation to new trajectories.