When facing complex decisions, people have been shown to seek advice in order to improve
their sub-optimal decision-making process. Advice-seeking and advice-provision is a fundamental practice in making real-world decisions. However, while automated systems are becoming increasingly prevalent in everyday life, the question of how an automated agent should
provide advice to a human user in order to enhance the user’s performance remains an open
challenge. This thesis studies the possibility of deploying intelligent automated agents to support and assist people in complex tasks and decisions through advice provision.
An automated advisor has a substantial computational advantage over its human user.
Therefore, if beneficially leveraged, this advantage can be translated into a significant enhancement in the human user’s performance in complex settings. There are many real-world settings
in which a human can benefit from the help of an intelligent advising agent. We study four
such settings in this thesis: 1) Argumentative dialogs where the user engages in a chat with
another human. The agent advises its user on which arguments the user should put forward
during the dialog and thus helps the user better persuade the other human or better present her
point of view. 2) Traffic enforcement where the user is a traffic police officer. The agent offers
advice to the officer as to where and when to place traffic enforcement measures in order to
minimize the expected number of traffic accidents. 3) Automotive Climate Control Systems
(CCSs) where the user is a human driver. The agent offers advice to the driver on how to set
her CSS in order to minimize the CCS’s energy consumption while keeping the driver comfortable. 4) Human-multi-robot team collaborations where the agent offers advice to a human
operator who operates a large team of semi-autonomous, low-cost robots. The agent seeks to
enhance the human-multi-robot team collaboration.
We propose and evaluate a novel methodology for designing and developing intelligent
advising agents. Our methodology heavily relies on the articulation and mitigation of complex optimization problems which involve people and, possibly, machines. According to our
methodology, an optimization problem is used to deriving the most beneficial advice provision
policy for an agent to deploy. The optimization problem accounts for both the user and the
environment which are modeled using machine learning and formal techniques. We provide
algorithms and heuristics for addressing this complex setting in four realistic real-world environments. Through extensive empirical evaluation, with more than a thousand human subjects,
electric vehicles and simulated and physical robots, we demonstrate the proposed methodology’s advantage over existing solutions. As part of our work on maximizing the advising
agents’ improvement, we contribute novel developments for several other sub-fields of Artificial Intelligence (AI).