Information-theoretic analysis of human-machine mixed systems

information theory, human-machine systems

Seo, Daewon

关键词: information theory, human-machine systems;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/105592/SEO-DISSERTATION-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Many recent information technologies such as crowdsourcing and social decision-making systems are designed based on (near-)optimal information processing techniques for machines. However, in such applications, some parts of systems that process information are humans and so systems are affected by bounded rationality of human behavior and overall performance is suboptimal. In this dissertation, we consider systems that include humans and study their information-theoretic limits. We investigate four problems in this direction and show fundamental limits in terms of capacity, Bayes risk, and rate-distortion.A system with queue-length-dependent service quality, motivated by crowdsourcing platforms, is investigated. Since human service quality changes depending on workload, a job designer must take the level of work into account. We model the workload using queueing theory and characterize Shannon's information capacity for single-user and multiuser systems.We also investigate social learning as sequential binary hypothesis testing. We find somewhat counterintuitively that unlike basic binary hypothesis testing, the decision threshold determined by the true prior probability is no longer optimal and biased perception of the true prior could outperform the unbiased perception system. The fact that the optimal belief curve resembles the Prelec weighting function from cumulative prospect theory gives insight, in the era of artificial intelligence (AI), into how to design machine AI that supports a human decision.The traditional CEO problem well models a collaborative decision-making problem. We extend the CEO problem to two continuous alphabet settings with general rth power of difference and logarithmic distortions, and study matching asymptotics of distortion as the number of agents and sum rate grow without bound.