Human-Allied Efficient and Effective Learning in Noisy Domains
AI agents must work in alliance with the humans, be it in mundane activities or in critical and strategic tasks to realize their full potential. This is crucial to envisage the objective of developing intelligent agents that can learn an act in noisy structured environment and are robust to observation density and quality. While extensive research and sophisticated techniques exist for faithfully modeling complex and rich information in structured noisy domains, robustness to density and quality are open challenges. Though AI agents are expected to learn/reason better with more data, yet arbitrarily large/dense data sets may, instead, lead to inefficiency, especially in structured domains. The problem is even more critical in sequential decision making, where the agent has to learn or reason continuously. The challenge of observation-quality, though distinct, is not completely orthogonal to the problem of density. Low-quality observations may originate due to systematic signal error, cognitive bias in data persistence as well as sample sparsity (known as systematic noise in general). Thus the two challenges are closely coupled. When the density becomes extremely low, it affects quality. Naive solutions for such interconnected and closely-coupled challenges, such as powerful computational hardware or explicit noise modeling with the help of gold-standard observations can be insufficient. This is due to the fact that the real problem lies at a deeper, more conceptual level and there is a need to develop domain independent solutions. For instance, it is difficult to model sysvii tematic noise, even reasonably well, using most existing noise modeling approaches and presence of gold-standard observations is a strong assumption. This dissertation investigates these challenges in depth and presents our unified Human-Allied AI (HAAI) framework to address them. It highlights some crucial insights in the context of both the challenges we address. Detailed algorithms are presented and analyzed as needed. We gain clarity about how Human-AI alliance, while not the more popular competitive or adversarial setting, is still a reasonable approach to address the above challenges. We also understand what a true HAAI system should entail and this work is a significant step in that direction. For instance progressive knowledge enhancement of the AI agent is one of the key aspects in HAAI, but is extremely difficult in real-time while the agent interacts with a human. One of subsequent chapters proposes principled approach to achieve the same. In summary this dissertation takes a significant step towards developing a robust intelligent agent. Finally, it outlines the open problems that are integral in developing true AI with common sense.