When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability
Extends Guided Soft Actor-Critic with Belief-Aware GSAC, which scales the privileged-distillation coefficient by ensemble disagreement in a driving POMDP. Across three occlusion levels, adaptive guidance helps under mild and moderate occlusion; under severe occlusion the coefficient collapses to its minimum early (“observability blindness”) and a simple linear decay schedule performs best, suggesting that much of the benefit comes from scheduling rather than adaptivity.
Research area: Reinforcement Learning & Autonomous Systems.
Dr. Mehmet Haklıdır is the founding Director of the TÜBİTAK BİLGEM Artificial Intelligence Institute, Türkiye’s first national AI institute. He holds a Ph.D. in Control and Automation Engineering from Istanbul Technical University and has more than 24 years of experience in AI, autonomous systems, big data, simulation, and technology policy.
He contributed to Türkiye’s first National AI Strategy and led the technical preparation and coordination of the National Artificial Intelligence Action Plan 2026–2030, finalized through consultation with the national AI ecosystem. He serves on the Steering Committee of the Science, Technology and Innovation Policies Council of the Presidency of the Republic of Türkiye and is a member of several international AI governance initiatives, including OECD ONE AI, GPAI, UNESCO AI Ethics, NATO DAIRB, and the European AI Alliance.
His current research focuses on large language models, AI safety with an emphasis on agentic AI systems, deep reinforcement learning, robotics, and autonomous decision-making under uncertainty.