AI in a vat: Fundamental limits of efficient world modelling for agent sandboxing and interpretability

Recent work proposes using world models to generate controlled virtual environments in which AI agents can be tested before deployment to ensure their reliability and safety. However, accurate world models often have high computational demands that can severely restrict the scope and depth of such assessments. Inspired by the classic `brain in a vat' thought experiment, here we investigate ways of simplifying world models that remain agnostic to the AI agent under evaluation. By following principles from computational mechanics, our approach reveals a fundamental trade-off in world model construction between efficiency and interpretability, demonstrating that no single world model can optimise all desirable characteristics. Building on this trade-off, we identify procedures to build world models that either minimise memory requirements, delineate the boundaries of what is learnable, or allow tracking causes of undesirable outcomes. In doing so, this work establishes fundamental limits in world modelling, leading to actionable guidelines that inform core design choices related to effective agent evaluation.

Mobile Cellular-Connected UAVs: Reinforcement Learning for Sky Limits

A cellular-connected unmanned aerial vehicle (UAV)faces several key challenges concerning connectivity and energy efficiency. Through a learning-based strategy, we propose a general novel multi-armed bandit (MAB) algorithm to reduce disconnectivity time, handover rate, and energy consumption of UAV by taking into account its time of task completion. By formulating the problem as a function of UAV's velocity, we show how each of these performance indicators (PIs) is improved by adopting a proper range of corresponding learning parameter, e.g. 50% reduction in HO rate as compared to a blind strategy. However, results reveal that the optimal combination of the learning parameters depends critically on any specific application and the weights of PIs on the final objective function.