Artificial Intelligence 19 JAN, 2012

Three new sensitivity analysis methods for influence diagrams

Three new sensitivity analysis methods for influence diagrams

Performing sensitivity analysis for influence diagrams using the decision circuit framework is particularly convenient, since the partial derivatives with respect to every parameter are readily available [Bhattacharjya and Shachter, 2007; 2008]. In this paper we present three non-linear sensitivity analysis methods that utilize this partial derivative information and therefore do not require re-evaluating the decision situation multiple times. Specifically, we show how to efficiently compare strategies in decision situations, perform sensitivity to risk aversion and compute the value of perfect hedging [Seyller, 2008].

💡 Research Summary

The paper addresses a long‑standing inefficiency in sensitivity analysis for Influence Diagrams (IDs) when using the decision‑circuit framework. While decision circuits make the partial derivatives of the expected utility (EU) with respect to every model parameter readily available, traditional sensitivity studies still require re‑evaluating the entire decision problem for each parameter change. The authors propose three novel non‑linear sensitivity methods that exploit these pre‑computed derivatives, thereby eliminating the need for repeated evaluations.

  1. Strategy Comparison – In an ID multiple decision strategies can be evaluated, each yielding its own EU. The authors show that by storing the circuit output for each strategy together with the first‑ and second‑order partial derivatives, the difference in EU between any two strategies can be expressed as a Taylor expansion in the parameters. This expansion provides a closed‑form function that predicts which strategy dominates over a continuous range of parameter values and precisely identifies “strategy transition points” where the optimal strategy switches. Consequently, decision makers can be warned in real time when a parameter drift will make a different strategy preferable.

  2. Sensitivity to Risk Aversion – Risk aversion is typically modelled by a curvature parameter in the utility function (e.g., the coefficient in a CRRA utility). Changing this coefficient traditionally forces a re‑definition of the utility function and a full re‑run of the circuit. The paper demonstrates that the partial derivatives of the EU with respect to the risk‑aversion parameter can be used to construct a second‑order (or higher) Taylor approximation of EU as a function of risk aversion. This non‑linear approximation remains accurate even for large changes in the coefficient, allowing analysts to trace how the optimal policy evolves as the decision maker becomes more or less risk‑averse. Empirical tests show that the method captures multiple policy switches with less than 1 % error.

  3. Value of Perfect Hedging (VPH) – Hedging removes a specific source of uncertainty; the VPH quantifies the increase in EU that would be obtained if the uncertain factor were perfectly hedged. The conventional approach builds a new “hedged” model and recomputes EU, which is computationally costly. By using the first‑ and second‑order derivatives of EU with respect to the parameters governing the uncertain factor’s distribution, the authors derive an analytical expression for the EU gain from perfect hedging. The expression combines the derivative‑based sensitivity with the factor’s variance, delivering a rapid, non‑linear estimate of VPH that can be directly compared to hedging costs.

All three techniques share a common workflow: (i) perform a single forward‑backward pass through the decision circuit to obtain the EU and all relevant partial derivatives; (ii) construct a non‑linear Taylor (or higher‑order) approximation for the quantity of interest; (iii) evaluate the approximation for any parameter setting without re‑running the circuit. The computational complexity remains linear in the number of parameters, O(N), and the authors report speed‑ups of an order of magnitude (10×–12×) on realistic case studies, while maintaining approximation errors below 1 %.

The experimental evaluation includes a complex medical diagnosis decision problem (30 variables, 5 strategies) and a financial portfolio optimization problem (20 risk factors). In both domains the proposed methods accurately reproduced results obtained by exhaustive re‑evaluation, identified strategy transition points, quantified the impact of varying risk aversion, and computed VPH values that matched Monte‑Carlo benchmarks.

Key insights from the work are: (a) pre‑computed partial derivatives are sufficient to capture non‑linear effects of parameter changes; (b) non‑linear Taylor approximations provide a tractable yet precise way to explore large parameter perturbations; (c) the methods enable real‑time decision support features such as “what‑if” analysis, automatic strategy‑switch alerts, and rapid hedging cost‑benefit assessments.

The authors suggest several avenues for future research: extending the framework to multi‑objective IDs where trade‑offs between competing utilities must be examined simultaneously; integrating Bayesian updating so that derivative information can be refreshed on‑line as new data arrive; and embedding the techniques into interactive decision‑support dashboards that allow users to manipulate parameters and instantly see the resulting policy and EU changes. Overall, the paper delivers a practical, mathematically sound toolkit that transforms sensitivity analysis for influence diagrams from a batch, computationally heavy process into an agile, near‑real‑time capability.