Finding Minimum Data Requirements Using Pseudo-independence
Yoonheui Kim, Victor Lesser · 2008
In situations where Bayesian networks (BN) inferencing approximation is allowable, we show how to reduce the amount of sensory observations necessary and in a multi-agent context the amount of agent communication. To achieve this, we introduce Pseudo-Independence, a relaxed independence relation that quantitatively differentiates the various degrees of independence among nodes in a BN. We combine Pseudo-Independence with Context-Specific Independence to obtain a measure, Context-Specific Pseudo-Independence (CSPI), that determines the amount of required data that needs to be used to infer within the error bound. We then use a Conditional Probability Table-based generation search process that utilize CSPI to determine the minimal observation set. We present empirical results to demonstrate that bounded approximate inference can be made with fewer observations.