Real-world problems are complex, uncertain, and changing. Problem solving and decision making in many critical areas , including building sustainable environments, managing disaster recovery, designing personalized medicine, engineering devices to help the aged, are often complicated by incomplete information, varying assumptions, and limited resources. These problems require solutions that are adaptive, usable, and cost-effective. These solutions mandate computational methodologies that can effectively manage "change".
Most AI techniques that manage complexity and uncertainty assume stationary environments; the underlying assumptions, world dynamics, action alternatives, and potential outcomes are predetermined and unchanged over the course of the problem solving process. In many real-world situations, however, the environment is non-stationary and perceptions are often different from expectations. In other words, the world model as the problem solver understands it is often changing. It is also impractical to pre-enumerate all the possible scenarios or alternatives beforehand.
- For a chronically ill patient with an acute complication, the physiological state of the patient may deteriorate over a period of time with different rates at different stages. How can a diagnostic test or treatment protocol advisor be adjusted accordingly?
- In a game situation, an aggressive player suddenly changes his tactics and starts to play defensively. How can a game AI detect the changes in the play pattern effectively and adjust countering strategies accordingly?
- In social networks and web-based interactions, some unseen emerging behavioral or communication patterns may be relevant to marketing agencies or intelligence departments. How can a monitoring program detect significant emerging patterns?
- For a robot roaming on unexplored Mars terrains, it might encounter dancing Martians who totally disable its original beliefs and plans. How can the robot adapt to the new surroundings and find the best course of actions?
Our current research builds on our past and on-going work in dynamic decision making in complex and uncertain domains with limited resources. Together with our research collaborators, we are working on a family of representation, reasoning, and learning techniques and frameworks that can effectively reflect and better manage the incomplete and changing nature of real-world problems.
The driving research hypothesis is as follows:
The world is non-stationary and computational solutions need to adapt to the varying environment factors and cognitive capabilities of the decision maker or problem solver.
Relevant problem representations and useful inferences should change so that new knowledge and insights can be learned more effectively in different situations.
On the other hand, smart learning techniques should detect changes in the environment, and automatically learn different models to effectively support future inferences.
Methodology and application foci
|Reasoning with Change||From Genes to Humans|