Risk-sensitive decision making under inclomplete information

点击翻译按钮获取中文摘要

Daily decisions have to be made in the face of risk that arises from incomplete information. For instance, a firefighter in a smoky burning house may be uncertain about whether she spotted a trapped victim and whether the rescue operation involves danger. Two different sources of uncertainty, from which risk arises, can be identified: (1) The incomplete knowledge of states of the world and (2) the uncertain consequences of future events at those states. We call them perceptual risk and economic risk. Both types of risk have to be taken into account by a decision maker.Decision making under risk has been a topic of research in a wide range of disciplines, such as finance, machine learning, control, operations research, behavioral economics, and cognitive neuroscience. However, both types of risk are usually investigated separately in two strands of research, and no integrated computational framework that incorporates both types of risk yet exists. The main goals of this research project are to develop an integrated computational framework for sequential decision making problems in face of both perceptual and economic risk and to derive computationally tractable algorithms to solve the corresponding optimization problems. The research is strongly driven by theory, and aims in its core at the extension of risk-sensitive reinforcement learning to partially observable Markov decision processes, which has never been done for a general setting.In order to test the applicability of the theoretical framework and the derived algorithms, we will evaluate both for optimal risk-sensitive decision making in stock market trading.