Lyapunov methods for safe intelligent agent design

摘要: In the many successful applications of artificial intelligence (AI) methods to real-world problems in domains such as medicine, commerce, and manufacturing, AI system usually plays an advisory or monitoring role. That is, provides information a human decision-maker, who has final say. However, for ranging from space exploration, e-commerce, search rescue missions, there is increasing need desire systems that display much greater degree autonomy. designing autonomous systems, agents, issues concerning safety, reliability, robustness become critical. Does agent observe appropriate safety constraints? Can we provide performance goal-achievement guarantees? deliberate and/or learn efficiently real time? In this dissertation, address some these by developing approach design integrates control-theoretic techniques, primarily based on Lyapunov functions, with planning learning techniques AI. Our main use domain knowledge formulate, restrict, ways which can interact its environment. This allows one construct agents enjoy provable guarantees, reason act real-time anytime fashion. Because guarantees are established restrictions agent's behavior, specialized “safety-oriented” decision-making algorithms not necessary. Agents using standard algorithms; discuss state-space reinforcement detail. To limited degree, also show needed ensure safe behavior itself be learned agent, known priori. We demonstrate our theory simulation experiments robotics control.