Çoklu görev öğrenimine insana benzer bir yaklaşım

This thesis proposes a novel human inspired framework for multi-task learning that challenges traditional, rigid training schedules. We design a system in which a learning agent autonomously selects which task to engage in based on internal, task-agnostic signals, by a high-level control mechanism, the Task Controller, which facilitates self-directed learning. We first develop LP-Con, a modular neural architecture for interleaved multi-task learning, where Learning Progress (LP) serves as the guiding signal. The model employs task-specific network blocks to enable flexible, many-to-many knowledge transfer and a human-like interleaved learning process. We then extend this with IMTL-EMLP, which integrates Energy Consumption (EC) into the Task Controller to enhance energy efficiency without sacrificing learning performance. Both models are evaluated in a simulated robotic environment where an agent learns to predict the consequences of its actions across various object interaction scenarios. To validate the effectiveness of our proposed model, we conduct several experiments which include overall task performance analysis, task-wise learning performances, task selection patterns by Task Controller, interleaving vs. blocked learning, and finally object-wise cross task skill transfer. The baselines include single-task learning, multi-task learning with random task scheduling, and blocked multi-task learning. Experimental results show that our cognitively inspired methods significantly accelerate learning, improve generalization, show robustness to forgetting and increase computational efficiency. By aligning learning strategies with the self-regulated, dynamic nature of human development, this work contributes to the creation of more adaptive, efficient, and scalable lifelong learning systems.