Consider the problem of monitoring the state of a complex dynamic system, and predicting its future evolution. Exact algorithms for this task typically maintain a belief state, or distribution over the states at some point in time. Unfortunately, these algorithms fail when applied to complex processes such as those represented as dynamic Bayesian networks (DBNs), as the representation of the belief state grows exponentially with the size of the process. In (Boyen & Koller 1998), we recently proposed an efficient approximate tracking algorithm that maintains an approximate belief state that has a compact representation as a set of independent factors. Its performance depends on the error introduced by approximating a belief state of this process by a factored one. We informally argued that this error is low if the interaction between variables in the processes is "weak". In this paper, we give formal information-theoretic definitions for notions such as weak interaction and sparse interaction of processes. We use these notions to analyze the conditions under which the error induced by this type of approximation is small. We demonstrate several cases where our results formally support intuitions about strength of interaction.
[1]
Keiji Kanazawa,et al.
A model for reasoning about persistence and causation
,
1989
.
[2]
Uffe Kjærulff,et al.
A Computational Scheme for Reasoning in Dynamic Probabilistic Networks
,
1992,
UAI.
[3]
Thomas M. Cover,et al.
Elements of Information Theory
,
2005
.
[4]
HERBERT A. SIMON,et al.
The Architecture of Complexity
,
1991
.
[5]
Avi Pfeffer,et al.
Structured Representation of Complex Stochastic Systems
,
1998,
AAAI/IAAI.
[6]
Karl Johan Åström,et al.
Optimal control of Markov processes with incomplete state information
,
1965
.
[7]
Stuart J. Russell,et al.
The BATmobile: Towards a Bayesian Automated Taxi
,
1995,
IJCAI.
[8]
Xavier Boyen,et al.
Tractable Inference for Complex Stochastic Processes
,
1998,
UAI.