Vadim Bulitko is interested in building the strong Artificial Intelligence as well as understanding intelligence and cognition in humans and animals. The following biography is organized by the area of his contributions.

AI: Real-Time Heuristic Search

Vadim has made the following contributions to the field of real-time heuristic search. First, he discovered lookahead pathology in single-agent real-time heuristic search (IJCAI, EMCRTS, ECAI, USC). Second, as a remedy to the pathology, Vadim suggested controlling lookahead dynamically via the use of pattern databases. This constituted the first application of pattern-databases to real-time heuristic search and the first use of pattern-databases to control search depth (AAAI workshop, ICAPS). The work was later extended by Vadim's colleagues to use decision trees instead of the pattern-databases as well as select goals dynamically (JAIR). Third, Vadim suggested a framework unifying and extending several well-known algorithms in real-time heuristic search (JAIR). Fourth, he proposed the first use of automatically built state abstraction in real-time heuristic search (AAAI, JAIR). In recognition of this work, Vadim was invited to give a talk at the Symposium on Abstraction, Reformulation and Approximation (SARA) in 2007. Additionally, applications of this work to pathfinding in video games have attracted attention of Bioware, Corp. Fifth, Vadim proposed using genetic algorithms with meta-models as a heuristic search in the space of MDP action sets (FEA, CEC, GECCO, GECCO, JMVLSC). Sixth, he suggested a hierarchical mini-max algorithm to trade-off coordination and planning speed in the domain of real-time multi-unit moving target pursuit. Vadim's current interests in the field lie with theoretical analysis of real-time heuristic search as well as with novel search applications such as to emotion modeling and cellular automaton evolution.

AI: Computer Games

In early 2006, with his M.Sc. student David Thue, Vadim started working on adaptive story generation for video games. The resulting system, called PaSSAGE, continuously monitors a player as he/she progresses through the game and maintains a player profile. The profile is used on-line to alter the story as it progresses. The work has been recognized by the academia (AIIDE 06, AIIDE 07, AAAI symposium) as well as by the industry in the form of a contribution to a book chapter in volume 4 of the renown AI Game Programming Wisdom book series. Since 2007, Vadim has been working with Marcia Spetch and her students on studying psychological mechanisms underlying search and hide behavior in simulated 3D environments. Knowing such patterns may help building better AI-controlled non-playable characters in video games. Since late 2006, together with Michael Bowling and students, Vadim has been involved in studying team-level human behavior in an on-line team first-person shooter, Counter-Strike. Vadim's students proposed the first method to describe team-level openings in the game as well as a Hidden Markov model for predicting enemy positions during the game. Most recently, Vadim has proposed a matrix-algebra-based approach to unifying emotion and culture modeling in virtual humans within virtual-reality training environments (presentation).

AI: Time Interval Petri Nets and Intelligent Tutoring

Vadim's research in Artificial  Intelligence started in January of 1996 with applications to shipboard damage control. His first project was development of a real-time decision-making system that coordinated damage control efforts aboard a naval vessel. First, Vadim developed a new flavor of Petri Nets suited for rapid modeling of concurrent damage spread processes. The formalism, called Time Interval Petri Nets (TIPN), supported temporal uncertainty of events and was used in the damage control decision system as a domain model for lookahead search (AIJ). Second, a decision-tree state evaluator was used to assess danger level of a ship state by predicting the amount of time until a major disaster. The decision tree was inferred from historic data on shipboard damage scenarios. In addition to outperforming human subjects in a large-scale simulated exercise by 318%, Vadim's damage control system was able to explain its behavior as well as evaluate courses of action taken by another damage control agent. This enabled its use as an intelligent tutoring tool (IAAI). The research was recognized by the AAAI in 1999 and received an Innovative Application of Artificial Intelligence award. This work represented the bulk of Vadim's M.Sc. thesis which was completed at the University of Illinois at Urbana-Champaign in May of 1998. Vadim then developed the first native algorithm for learning Time Interval Petri Nets and compared its performance to inductive logic programming methods (LNAI, JMVLSC). He received a Ph.D. for this work from the University of Illinois at Urbana-Champaign in 1999 under the supervision of David C. Wilkins.

AI: Reinforcement Learning

Vadim's contributions to RL are three-fold. First, he suggested a use of predictive state representations as synthetic sensors for AI agents in video games (IJCAI). The work allowed one to augment agent's built-in sensors with high-level predictions (e.g., "against a wall", "in a corner", "in a narrow corridor") useful for in-game AI-controlled tactical behavior. Second, Vadim's students and he proposed an application of prioritized sweeping to real-time heuristic search and empirically demonstrated its potential in real-time pathfinding (IJCAI). Third, together with his student and Russ Greiner, Vadim has worked on focused reinforcement learning wherein an agent attempts to learn about the most important areas of the state-action space first (J.UCS).

AI: Computer Vision

Traditional deployed computer vision systems are primarily manually engineered. Their performance thus depends on years of human effort and substantial computer vision and forestry expertise. In 2001, Vadim took an alternative approach and treated the problem as a control task over a large library of readily available computer vision routines. Together with his students and in collaboration with the Alberta Research Council, he successfully applied machine learning methods to derive such a control policy automatically from training data. Namely, a then state-of-the-art approach for machine learnable vision systems (Bruce Draper’s ADORE) was adopted with the intention of removing the last vestiges of human intervention from it. The advances included a use of domain-independent off-the-shelf vision libraries, least-commitment policies with reduced needs for manually engineered state features, learning from partially labeled data, automated feature selection and automated selection of compact high-performance sub-libraries of vision operators. An implementation of this approach was successfully trained to recognize tree canopies in aerial photographs - a challenging task with 18 - 40% of human expert error. This was the first demonstration of such an approach successfully applied to recognition of natural (as opposed to man-made) objects (IAAI, NASA). Later, the system was re-trained to label chunks of oil sand ore in a joint project with the University of Alberta Centre for Intelligent Mining Systems and Syncrude Research, Ltd. Part of this work was published as a contribution to a book chapter on feature selection.

Recursion Theory

In the early 1990s, Vadim worked in recursion theory and proposed a new completeness criterion for pseudo-simple sets (J.UCS). He has continued his work in the field, publishing several more journal papers (UMJ, MLQ). Vadim received his B.Sc. in mathematics with a minor in psychology from Odessa State University in the summer of 1995.

Software Engineering

Vadim started programming and tinkering with computers at the age of eight using programmable calculators and later PDP-11 clones. His early contributions were an implementation of a relationship database system similar to dBase-III+ on a machine with 16Kbytes of RAM and no hard drive as well as writing a thread scheduler. In 1994, Vadim worked as a software developer in a Silicon Valley computer vision company, programming for x86 hardware as well as AT&T DSP boards.

Service

Vadim is a co-chair of SARA'09. In 2006, Vadim was a tutorial and workshop chair for ICML. In 2005, he and Sven Koenig co-chaired an IJCAI workshop on search in uncertain and dynamic environments. In the summer of 2002, Vadim organized, chaired and taught at a Summer School on Quantum Computing. He is a regular reviewer for JAIR, AAAI, IJCAI, ICML and other top AI journals and conferences.

Teaching

Since 2000, Vadim has taught and coordinated courses at the University of Alberta at graduate and undergraduate levels. The course materials that Vadim designed have since been used by several other teaching institutions. He has received favorable feedback from his students and, as a result, was recognized by the Faculty of Science dean as a "top echelon teacher" in 2006 and 2007. Since late 2007 Vadim has been serving on the undergraduate curriculum committee at the Computing Science department, University of Alberta.

Supervision

Vadim started supervising as the head of his research group, Intelligent Reasoning, Critiquing and Learning (IRCL), which was formed in January 1996 as a part of the knowledge-based systems group at the University of Illinois at Urbana-Champaign. Since then, Vadim has supervised and co-supervised 40 graduate, undergraduate and project students.