Abstract : Probabilistic Graphical Models (PGMs) provide a normative tool for modeling uncertain, dynamic situations and at different levels of detail. These models, which include Bayesian and Markov networks, allow users to model the causal influences in a situation qualitatively using the language of graphs, and also quantitatively using probabilities and compatibilities (called model parameters). This quantitative aspect of PGMs is their strongest and weakest point. Its strength comes from providing an ability to model situations at a fine grained level, yet experts may find these numeric parameters counter intuitive to specify, interpret and control. Sensitivity analysis is concerned with characterizing the relationship between the answers to model queries and the values of model parameters, allowing decision makers to get first hand insights into the sensitivity and robustness of their decisions to the various assumptions underlying the situation model (as exhibited by model parameters). This report describes some key results obtained on the theory and practice of sensitivity analysis in probabilistic graphical models.
[1]
Adnan Darwiche,et al.
Node Splitting: A Scheme for Generating Upper Bounds in Bayesian Networks
,
2007,
UAI.
[2]
Adnan Darwiche,et al.
On Bayesian Network Approximation by Edge Deletion
,
2005,
UAI.
[3]
Adnan Darwiche,et al.
On the Robustness of Most Probable Explanations
,
2006,
UAI.
[4]
Adnan Darwiche,et al.
A Variational Approach for Approximating Bayesian Networks by Edge Deletion
,
2006,
UAI.
[5]
Adnan Darwiche,et al.
Compiling Bayesian Networks with Local Structure
,
2005,
IJCAI.
[6]
Adnan Darwiche,et al.
On probabilistic inference by weighted model counting
,
2008,
Artif. Intell..
[7]
Adnan Darwiche,et al.
Sensitivity Analysis in Markov Networks
,
2005,
IJCAI.
[8]
Adnan Darwiche,et al.
An Edge Deletion Semantics for Belief Propagation and its Practical Impact on Approximation Quality
,
2006,
AAAI.