PED: A Planner for Efficient Dialogues

Bryan McEleney PhD

Department of Computer Science, University College Dublin.

Imagine this problem - you are 90 percent sure that it is not raining outside. Would a dialogue asking whether it is raining be a productive investment of your time, or should you just take a chance and go outside? This type of problem is very common where we must decide which subject or direction in a dialogue is the most valuable. Until now there has been no general dialogue planning system for this important problem.



PED is a dialogue management system that uses a probabilistic nested belief model to choose dialogue strategies from a Bayesian game tree. A nested belief model is a representation of the beliefs of the speakers in the dialogue, including their beliefs about each other. PED constructs game trees to model the possible outcomes of the dialogue. In the example tree depicted above, there are nodes to represent decision options ( circles ) and nodes to represent chance outcomes ( squares ). The game tree is evaluated in the context of the belief model to estimate the chance outcomes and determine the next move in the dialogue. PED uses a belief revision function that makes inferences from the moves observed in the dialogue to update the belief model as the dialogue progresses.

PED was developed as my PhD work at University College Dublin and was funded by the Irish government.

To browse my thesis on the web, click here.

To download the PDF version, click here.

For something shorter, try one of these papers:

Bryan McEleney, Gregory O'Hare
Decision Theoretic Planning for Initiative Problems
Proceedings 10th International Conference on User Modelling 2005
July 23rd-29th 2005, Edinburgh, Scotland.

Bryan McEleney, Gregory O'Hare
Efficient Dialogue Using a Probabilistic Nested User Model
Proceedings of the Fourth IJCAI workshop on Knowledge and Reasoning in Practical Dialogue Systems
August 1st 2005, Edinburgh, Scotland.


To download a package with the PED source code, experiment scripts and experiment results, click here. The SWI Prolog interpreter using GNU/Linux is the recommended environment for running PED.

Bryan McEleney, December 2006.