16 January 2018 in Nature Communications
Since Alan Turing envisioned artificial intelligence, technical progress has often been measured by the ability to defeat humans in zero-sum encounters (e.g., Chess, Poker, or Go). Less attention has been given to scenarios in which human–machine cooperation is beneficial but non-trivial, such as scenarios in which human and machine preferences are neither fully aligned nor fully in conflict. Cooperation does not require sheer computational power, but instead is facilitated by intuition, cultural norms, emotions, signals, and pre-evolved dispositions. Here, we develop an algorithm that combines a state-of-the-art reinforcement-learning algorithm with mechanisms for signaling. We show that this algorithm can cooperate with people and other algorithms at levels that rival human cooperation in a variety of two-player repeated stochastic games. These results indicate that general human–machine cooperation is achievable using a non-trivial, but ultimately simple, set of algorithmic mechanisms.
12 November 2019 in Nature Machine Intelligence
12 April 2018 in IEEE Transactions on Smart Grid
Our ever-increasing reliance on electricity coupled with inefficient consumption has resulted in several economical and environmental threats. To curb these threats, smart grids are emerging. These improved power systems could potentially reduce the peak consumption and better match demand to supply, to produce both economical and environmental advantages. In this work, we consider two pricing problems in the smart grid, under a dynamic pricing model, where consumers follow threshold policies to schedule their power consumption. The first problem is to set the prices during the different time periods such that the peak demand is minimized. The second problem is to set the prices such that the power demand matches the supply. Firstly, we propose generic heuristics called GREEDY and SLIDING-WINDOW that are able to solve the two studied problems in addition to any other optimization problem, under the same model. Secondly, we provide theoretical analysis for the uniform-pricing approach in the context of peak-demand minimization. In addition, we propose optimal algorithms for the two optimization problems that can be used when the maximum deadline period of the power jobs is relatively small. Moreover, we conduct several experiments to evaluate the proposed algorithms and the uniform pricing approach on real data. Our experimental results showed that our proposed heuristics have a relatively low approximation ratio, and have the potential to provide a significant energy saving in many cases compared to the Time-of-Use (ToU) pricing. Furthermore, the experiments showed that while the uniform pricing has an acceptable approximation ratio in the average case, it leads to energy loss compared to the ToU pricing. Finally, the experiments demonstrated a tradeoff between optimality and speed.
7 April 2020 in Proceedings of the National Academy of Sciences
20 April 2023 in IEEE Transactions on Robotics
While the design of autonomous robots often emphasizes developing proficient robots, another important attribute of autonomous robot systems is their ability to evaluate their own proficiency. A robot should be able to assess how well it can perform a task before, during, and after it has attempted the task. How can autonomous robots be designed to self-assess their behavior? This article presents the assumption-alignment tracking (AAT) method for designing autonomous robots that can effectively evaluate their own performance. In AAT, the robot a) tracks the veracity of assumptions made by the robot's decision-making algorithms to measure how well these algorithms fit, or align with , its environment and hardware systems, and b) uses the measurement of alignment to assess the robot's ability to succeed at a given task based on its past experiences. The efficacy of AAT is illustrated through three case studies: a simulated robot navigating in a maze-based (discrete time) Markov chain environment, a simulated robot navigating in a continuous environment, and a real-world robot arranging blocks of different shapes and colors in a specific order on a table. Results show that AAT is able to accurately predict robot performance and, hence, determine robot proficiency in real time.