Information/Relevance Influence Diagrams

TL;DR

This paper introduces information/relevance influence diagrams (IRIDs), an extension of influence diagrams that directly represent decision constraints, solved using stochastic dynamic programming and Gibbs sampling, improving handling of complex decision problems.

Contribution

The paper proposes IRIDs, enabling direct representation of decision constraints within influence diagrams, and introduces a combined stochastic dynamic programming and Gibbs sampling approach for solving them.

Findings

IRIDs allow explicit modeling of decision constraints.

The combined solution method handles complex IRIDs effectively.

IRIDs improve decision modeling flexibility.

Abstract

In this paper we extend the influence diagram (ID) representation for decisions under uncertainty. In the standard ID, arrows into a decision node are only informational; they do not represent constraints on what the decision maker can do. We can represent such constraints only indirectly, using arrows to the children of the decision and sometimes adding more variables to the influence diagram, thus making the ID more complicated. Users of influence diagrams often want to represent constraints by arrows into decision nodes. We represent constraints on decisions by allowing relevance arrows into decision nodes. We call the resulting representation information/relevance influence diagrams (IRIDs). Information/relevance influence diagrams allow for direct representation and specification of constrained decisions. We use a combination of stochastic dynamic programming and Gibbs sampling to solve IRIDs. This method is especially useful when exact methods for solving IDs fail.