Context-Aware Artificial Intelligence Systems for Long-Horizon Decision Making
Abstract
Long-horizon decision making remains a fundamental challenge in artificial intelligence, particularly in complex, dynamic environments where agents must reason over extended temporal sequences while maintaining contextual awareness. This paper presents a novel Context-Aware Long-Horizon Decision Making (CA-LHDM) framework that integrates hierarchical reinforcement learning with transformer-based context encoding to enable effective planning and execution over extended time horizons. Our approach leverages attention mechanisms to capture temporal dependencies and contextual information across multiple abstraction levels, enabling agents to decompose complex tasks into manageable subgoals while maintaining global coherence. We introduce a hybrid architecture combining model-based planning with model-free learning, supported by an adaptive memory module that selectively retains relevant historical information. Experimental evaluation on benchmark tasks including MuJoCo continuous control and multi-agent coordination scenarios demonstrates that CA-LHDM achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 7 {\%}}$</tex> improvement in cumulative reward and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 5 {\%}}$</tex> reduction in planning time compared to state-of-the-art baselines. Our ablation studies reveal the critical role of context encoding in achieving superior performance, with the attention-based context module contributing 28% of the overall performance gain.