I am interested in the acquisition, representation, organization and retrieval of knowledge and its use in reasoning.
My dissertation research investigates generic ways for an intelligent agent to organize and retrieve its past experience in an accurate and scalable way.
I have developed a generic episodic memory module that can be attached to a variety of applications and used for different tasks. Encapsulating the complexity of such a memory into a separate subsystem should reduce the complexity the overall system while allowing research to focus on the generic aspects of memory organization and retrieval in isolation of a specific domain and task.
An intelligent system can thus can store, recall and reuse experience of virtually every reasoning action taken. This is a different kind of knowledge that captures the result and also the context in which a reasoning action was taken. Availability of such knowledge can improve both the correctness and performance of a reasoning system by focusing on actions that are more likely to produce desirable results, given the history of their past applications.
A full description of the design, implementation and evaluation of the memory module can be found in my dissertation, entitled "A Generic Memory Module for Events".
A shorter description of memory module can be found in an earlier paper, presented at FLAIRS 20.
The generic memory module was applied to different tasks such as:
- Planning - this is our initial study of the memory indexing mechanism; an evaluation on a dataset in the Logistics domain showed that:
- indexing was able to reduce the search effort, while preserving accuracy when compared to a kNN algorithm.
- search effort did not grow as fast as the memory size
- same memory structure can be used for other tasks like classification and plan recognition.
- Plan recognition - an episodic-based approach to plan recognition is proposed; it is able to grow the plan library and make incremental predictions; we evaluated in on two datasets (Linux and Monroe); results show that:
- the episodic-based approach achieves similar performance as a statistical approach (Blaylock and Allen 05) on the goal-schema recognition task, on both corpora, while converging on more sessions.
- parameter recognition performance is poorer; we think this might be due to two things: the way an overall mapping from a new plan to a prior plan is constructed and to the lack of an adaptation strategy.
- Retrieval effort does not grow at the same rate as memory size.
A publication is forthcoming, but See the AAAI Fall Symposium 06 paper and the FLAIRS 22 paper for applications of episodic memory module to goal schema recognition (a plan recognition subtask).
- Physics problem-solving - the episodic memory is used to augment a problem-solver that systematically searches a KB for the most appropriate model that can answer and explain AP-level questions in the Physics domain. Memory is used to make informed suggestions (based on past experience) on what models should be used by the problem solver. Experiments show that:
- Memory is able to significantly reduce problem solving time when compared to a memory-less problem solver; the change is even larger when multiple models need to be applied to solve a question.
- Memory overhead represents only a fraction of the problem solving time.
- Memory overhead does not grow at the same rate as memory size
A publication is forthcoming.
Last modified: Thu Aug 02 11:16:31 Eastern Daylight Time 2012