Search engines and recommender systems are the primary tools available today to assist people in overcoming information overload. While they are very useful, they can only help users access relevant information but cannot support a user's task. In general, a user needs to find relevant information/knowledge in order to finish a task, thus information access is often a means to the end of finishing a task. To apply AI techniques to support a user's task, we are interested in developing intelligent task agents, especially general models and algorithms that can be used to build specific intelligent task agents in many different application domains.
An intelligent agent is "an agent acting in an intelligent manner; It perceives its environment, takes actions autonomously in order to achieve goals, and may improve its performance with learning or acquiring knowledge." (from Wikipedia). An intelligent task agent (ITA) is an intelligent agent whose goal is to help a user finish a task with no/minimum effort from the user.
We are especially interested in developing ITAs that can help users with complex tasks that involve the use of information or knowledge for decision making in an interactive way to optimize the collaboration of users and AI techniques. As such, we are interested in developing the ITAs that can leverage big data analysis to acquire knowledge from the data (see the DataScope Project) as well as personalize their interactions with a user based on formal models of users (see the User simulation Project )
Our current work includes the following directions:
- Formal frameworks and models for ITA: We are interested in developing mathematical models of ITAs and using them to guide us in designing and implementing general algorithms for supporting many specific ITAs in different application domains. A specific vision is the Intelligent Interactive Information Agent (I3A) model presented in the keynote talk at ACM ICTIR 2022. This is based on the game-theoretic framework for information retrieval, where the interaction of an agent and a user is framed a cooperative game and the agent's actions are optimized via a Bayesian decision framework. The framework is an extension of a previous risk minimization framework for information retrieval.
- Interface Computing: In most computational applications, the algorithms are responsible for generating the desired output from the given input, but the design of user interface is generally designed manually (heuristically). We are interested in studying how to use an algorithm to automatically generate an optimal user interface, which we refer to as "interface computing". As user interaction becomes increasingly complicated, how to specify the input and how to show the output from an algorithm have also become increasingly complex, leading to complex multimedia interfaces that we see in many online information and knowlege service systems. As we envision ITAs to add more interaction and conversational functions to such complex information service systems, the interface will likely become even more complex. We are interested in studying how to develop intelligent algorithms to automatically generate an optimal interaction interface that can adapt to a specific user in a specific interaction context with consideration of the size of the screen for displaying an interface. See our work on the Interface Card Model (ACM SIGIR'15 paper, >ACM SIGIR'16 paper), where we have shown that it's possible to use an algorithm to generate a browsing interface adapted to the screen size of the device (e.g., a mobile phone vs. a laptop) as well as to the system's confidence in the inferred user information need. Also see a talk on Interface Computing.
- Social Chatbot: The CharmBana Team of UIUC participated in the Alexa Socialbot Grand Challenge 5.
- Domain-Specific ITAs:
