|
My research aims to reduce the difficulty and cost of building AI systems, enable users who have no programming background to create or customize them, and equip AI systems with the ability to improve themselves by interacting with humans. My research combines ideas from artificial intelligence, machine learning, and cognitive science and I often apply these ideas to computer games and simulated environments. Intelligence often involves performing tasks by selecting actions over time. I investigate how a system can learn task-performance knowledge by observing other agents, interacting with them, and revising acquired knowledge by trying them in performing tasks (see the video excerpt below motivating several of these research directions). I am also interested in a related topic, transfer learning, which involves systems that reuse knowledge from previous experiences in distinct but related tasks.
In this video, Lieutenant Commander Data learning complex dance skills (Star Trek TNG Ep. 4.11 © Paramount Inc.). Mr. Data acquires new skills by observing and interacting with an expert and then he refines those skills with further experimentation. I often apply my research to Computer Games and Simulated Environments because they provide complex testbeds. Digital Entertainment is a big and growing industry that requires a rapid software development cycle and close interaction with artists who might not have programming skills. Both of these requirements makes Digital Entertainment an excellent area to apply my research on (semi-)automated generation of intelligent software agents. I conduct machine learning research with a focus on using multiple rich knowledge sources as input, such as logical rules, databases, or relational temporal data. These problems often require a combination of ideas from relational machine learning (i.e. Inductive Logic Programming) and statistical machine learning. Although most systems I investigate use considerable domain knowledge, I am looking for domain independent machine learning and performance system solutions. Therefore, my research often relies on cognitive architectures that make a clear distinction between domain dependent knowledge and domain independent architecture. Below is a summary of my work in these and other topics. Most of these works are done with close collaboration with many colleagues. Research Interests
Publications |
