---
title: Conditional Random-Utility Representation
url: https://www.emergentmind.com/topics/conditional-random-utility-representation
type: topic
---

# Conditional Random-Utility Representation

Conditional random-utility representation denotes a family of decision-theoretic and choice-theoretic formalisms in which utility, random utility, or expected utility is specified through conditional objects and then assembled into a global evaluative criterion. In the literature considered here, conditioning appears in several distinct but related forms: conditional utility measures and utility independence in utility distributions and utility networks [1302.1568]; ceteris-paribus utility ratios and conditional expected utility independence in expected utility networks [1301.6714]; multilinear utility factorizations in directed expected utility networks [1608.00810]; world- and observation-conditional plan utility in situation-calculus planning with Independent Choice Logic [1302.3597]; history-conditional stochastic utility in dynamic random choice [2102.00143]; and neural approximations of random utility models over product features, customer features, and assortments in RUMnets [2207.12877]. The common thread is that utility-side structure is treated as conditionable, modular, and computationally exploitable in close analogy to probabilistic structure.

## 1. Conditional utility as a measure-theoretic analogue of probability

A foundational route to conditional random-utility representation begins with the reinterpretation of utility in explicitly probabilistic terms. In Shoham’s framework, multiattribute utility theory starts from attributes \(X_1,\dots,X_n\), with a general multiattribute utility function \(U(x_1,\dots,x_n)\). Under the usual additive-independence axiom, there exist univariate functions \(u_i\) and constants \(k_i\) such that
\[
U(x_1,\dots,x_n)=\sum_{i=1}^n k_i u_i(x_i).
\]
In the special Boolean “Take-It-Or-Leave-It” case, \(x_i\in\{0,1\}\), \(u_i(1)=1\), \(u_i(0)=0\), \(k_i\ge 0\), and \(\sum_i k_i=1\), yielding the utility distribution
\[
U(x_1,\dots,x_n)=\sum_{i=1}^n k_i x_i.
\]
This extends to subsets \(S\subseteq\{1,\dots,n\}\) by finite additivity:
\[
U(S)=\sum_{i\in S} k_i.
\]
On that basis, conditional utility is defined exactly as conditional probability:
\[
U(A\mid B)\;:=\;\frac{U(A\cap B)}{U(B)},\qquad U(B)>0.
\]
The familiar identities then carry over: the chain rule \(U(A\cap B)=U(B)\cdot U(A\mid B)\), normalization \(U(\Omega\mid B)=1\), and a law of total utility [1302.1568].

Utility independence is likewise defined in direct parallel with probabilistic independence:
\[
U(A\cap C\mid B)=U(A\mid B)\cdot U(C\mid B).
\]
A utility network, or \(u\)-net, is then a directed acyclic graph whose local Markov property is stated under \(\perp_u\), so that the joint utility distribution factorizes into local conditional-utility tables,
\[
U(x_1,\dots,x_n)=\prod_{i=1}^n U(x_i\mid \mathrm{Pa}(x_i)).
\]
The explicit claim in this framework is that a \(u\)-net does for utilities what a Bayesian network does for probabilities, and that the resulting representation is “precisely a conditional random-utility model in which local conditional-utility tables replace local CP-tables” [1302.1568].

This interpretation also introduces several important caveats. The factors need not correspond to complete states of the world but may instead be atomic “teleological” ingredients of utility. Conditional utilities \(U(A\mid B)\) can exceed \(1\), because utility mass is renormalized given \(B\). Independence is interpreted teleologically or motivationally, not causally [1302.1568]. These departures matter because they distinguish conditional utility representations from both standard von Neumann–Morgenstern state utility and ordinary probabilistic conditioning.

## 2. Graphical representations of conditional utility and expected utility

A second line of development makes the conditioning structure explicitly graphical. In the expected utility network (EUN) of La Mura and Shoham, one works with finite random variables \(X_1,\dots,X_n\) and a reference state \(x^0\). An EUN is an undirected graph
\[
G=(N,E_p\cup E_u),
\]
with two disjoint edge sets: probability arcs \(E_p\) and utility arcs \(E_u\). Each node \(i\) carries two positive potentials,
\[
q_i(x_i\mid x_{P(i)})\qquad\text{and}\qquad w_i(x_i\mid x_{U(i)}),
\]
interpreted as ceteris-paribus probability ratios and utility ratios relative to the reference state. The global factorization is
\[
p(x)=p(x^0)\prod_{i=1}^n q_i(x_i\mid x_{P(i)}),
\]
and
\[
\frac{u(x)}{u(x^0)}=\prod_{i=1}^n w_i(x_i\mid x_{U(i)}).
\]
Thus, probability and utility are both modular, but they are modularized by different subgraphs [1301.6714].

Within this representation, utility independence is not formulated additively. Rather, for \(X_M\) and conditioning set \(X_K\),
\[
X_M\perp_u X_{N\setminus(M\cup K)}\mid X_K
\]
holds when the multiplicative utility ratio satisfies
\[
w(x_M\mid x_{N\setminus M})=w(x_M\mid x_K)\qquad \forall x_N.
\]
This differs sharply from probabilistic conditional independence, which constrains conditional probabilities rather than utility ratios [1301.6714].

EUNs also define conditional expected utility at the level of events. Utility is extended from states to events \(E\) by
\[
u(E)=\frac{\sum_{x\in E}u(x)\,p(x)}{p(E)},
\]
with \(u(\mathrm{True})=1\). Two events \(E,F\) are conditionally EU-independent given \(G\) when
\[
u(E\cap F\mid G)=u(E\mid G)\,u(F\mid G).
\]
A key theorem states that if \(C\) separates \(A\) from \(B\) in both the probability and utility subgraphs, then \(A\) and \(B\) are conditionally EU-independent given \(C\) [1301.6714]. This result links structural separation directly to strategic inference.

The conceptual significance is that conditional random-utility representation is no longer merely a way of storing local utility tables. It becomes a joint graph-theoretic language for stating when expected utility computations can be localized, simplified, or decomposed.

## 3. Directed expected utility networks and multilinear factorization

Leonelli and Smith’s directed expected utility network (DEUN) extends this agenda by combining a Bayesian-network factorization for probability with a directional utility diagram for utility. A DEUN on attributes \(Y_1,\dots,Y_n\) has two kinds of directed edges:
\[
(V(\mathcal G),E_p(\mathcal G))\quad\text{and}\quad (V(\mathcal G),E_u(\mathcal G)),
\]
where solid arrows define a BN and dashed arrows define a directional utility diagram, with the ordering restriction \(i<j\) for every edge \(i\to j\). The probabilistic factorization is
\[
p(\mathbf y)=\prod_{i=1}^n p(y_i\mid \mathbf y_{\Pi_i^p}),
\]
while the utility factorization is multilinear:
\[
u(\mathbf y)=\sum_{\mathbf y^{0*}\in \mathcal Y^{0*}}u(\mathbf y^{0*})\prod_{i=1}^n g(y_i\mid \mathbf y_{\Pi_i^u}^{0*}),
\]
where \(g(y_i\mid\cdot)\) is either \(u(y_i\mid\cdot)\) or \(1-u(y_i\mid\cdot)\) [1608.00810].

The absence of a dashed edge \(i\to j\) encodes utility conditional independence:
\[
Y_j\;\text{UI}\;Y_i\mid \mathbf Y_{-\{i,j\}},
\]
that is, \(u(y_j\mid y_i,y_{-ij})=u(y_j\mid y_{-ij})\). When the dashed-arrow graph is acyclic, the utility function admits the multilinear expansion above [1608.00810].

Under Savage’s expected-utility rule, a deterministic decision \(d\) has
\[
EU(d)=\sum_{\mathbf y\in\mathcal Y}p(\mathbf y\mid d)\,u(\mathbf y,d).
\]
In the DEUN treatment, the explicit \(d\) may be suppressed when the graphical structure is fixed across decisions, yielding
\[
EU=\sum_{\mathbf y}p(\mathbf y)\,u(\mathbf y).
\]
The practical importance lies in the fact that the double factorization of \(p\) and \(u\) permits a distributed evaluation of expected utility by successive one-dimensional integrals [1608.00810].

The generic backward-induction algorithm defines local utility vectors \(\mathbf u_i(y_i\mid \mathbf y_{\Pi_i^u}^{0*})\), combines them by the element-wise matching product \(\circ\), and computes
\[
\overline{\mathbf u_n}
=\int \mathbf u_n(y_n\mid \mathbf y_{\Pi_n^u}^{0*})\,p(y_n\mid \mathbf y_{\Pi_n^p})\,dy_n,
\]
then recursively for \(i=n-1,\dots,1\),
\[
\overline{\mathbf u_i}
=\int \Bigl(\overline{\mathbf u_{i+1}}\circ \mathbf u_i(y_i\mid \mathbf y_{\Pi_i^u}^{0*})\Bigr)\,p(y_i\mid \mathbf y_{\Pi_i^p})\,dy_i,
\]
and finally
\[
EU=\Bigl|\mathbf u_0(\mathbf y^{0*})\circ \overline{\mathbf u_1}\Bigr|.
\]
For decomposable DEUNs, the evaluation can instead be run on a standard junction tree using clique probability potentials \(\Phi_C\) and utility potentials \(\Psi_C\), with leaf absorption defined by
\[
\Psi_P\longmapsto \Psi_P\circ \int (\Psi_L\,\Phi_L)\,d\mathbf y_{L\setminus S_L}.
\]
After a full backward sweep,
\[
EU=\Bigl|\int \Phi(\mathbf y_C)\,\Psi(\mathbf y_C)\,d\mathbf y_C\Bigr|.
\]
These routines are presented as generalizations of standard influence-diagram evaluation to multilinear utility factorizations [1608.00810].

The food-security example in the same framework shows how a DEUN can support a four-node policy analysis with attributes \(Y_H\), \(Y_E\), \(Y_S\), and \(Y_C\), three decisions \(d_0,d_1,d_2\), Normal regression models for the attributes, and exponential-type conditional utilities normalized into the unit interval. With one realistic parameter choice reported in the paper, the expected utilities are approximately \(EU(d_0)\approx 0.29\), \(EU(d_1)\approx 0.19\), and \(EU(d_2)\approx 0.21\), so increasing the free-meal eligibility is optimal [1608.00810].

## 4. Conditional plans, stochastic frame axioms, and random utility in logic

Poole’s framework places conditional random-utility representation inside a logical action formalism. The setting is McCarthy’s situation calculus, with situations built from the initial situation \(s_0\) and successor situations \(do(a,S)\) when the agent attempts primitive action \(a\) in situation \(S\). Fluents are predicates \(F(\ldots,S)\) whose truth at \(S\) depends on both the last action and stochastic mechanisms represented by atomic choices. A positive-effect axiom may take the form
\[
carrying(key,do(pickup(key),S))
\leftarrow at(robot,Pos,S)\wedge at(key,Pos,S)\wedge pickup\_succeeds(S),
\]
while a frame axiom may take the form
\[
\begin{aligned}
carrying(key,do(A,S)) \leftarrow\;& carrying(key,S)\;\wedge\;A\neq putdown(key)\wedge A\neq pickup(key)\\
&\wedge\; keeps\_carrying(key,S).
\end{aligned}
\]
By making \(pickup\_succeeds(S)\) and \(keeps\_carrying(key,S)\) atomic choices, the formalism obtains stochastic frame axioms [1302.3597].

The uncertainty model is an Independent Choice Logic theory
\[
(\mathcal C,A,O,P_o,F),
\]
where \(\mathcal C\) is a choice space of disjoint alternatives of atomic choices, \(A\) is the set of primitive actions, \(O\) is the set of observable sensor-values, \(P_o\) assigns a probability distribution over each alternative, and \(F\) is an acyclic logic program whose unique Clark-completion model yields the non-stochastic consequences of every total choice. A possible world \(w_\tau\) is induced by a selector \(\tau\) choosing one atom from each \(C\in\mathcal C\), with probability
\[
P(w_\tau)=\prod_{C\in\mathcal C} P_o(\tau(C)).
\]
Utility is represented by a fluent-style predicate \(\mathrm{utility}(U,S)\) in \(F\), with the utility-completeness condition that for every world \(w\) and situation \(S\) there is exactly one real \(U\) such that \(w\models \mathrm{utility}(U,S)\) [1302.3597].

This permits arbitrary functions of final states to be encoded through fluent rules. One of Poole’s examples writes final utility as “remaining resources \(R\) plus prize \(P\)”:
\[
utility(R+P,S)\leftarrow resources(R,S)\wedge prize(P,S),
\]
with prize determined by whether the robot crashed, reached the lab, or neither, and resources updated by successor-state axioms [1302.3597].

A conditional plan \(P\) is generated by
\[
P:=skip\mid a\mid P;Q\mid (\mathrm{if}\;C\;\mathrm{then}\;P\;\mathrm{else}\;Q\;\mathrm{endIf}),
\]
where \(a\in A\) and \(C\in O\). Its semantics are given by a transition relation \(\mathit{trans}(P,w,S_1,S_2)\), so that branching depends on whether \(w\models sense(C,S_1)\). In each world \(w\), exactly one final situation \(S\) arises from \(P\), and the induced utility is
\[
u(w,P)=U
\quad\text{iff}\quad
\mathit{trans}(P,w,s_0,S)\;\text{and}\; w\models \mathrm{utility}(U,S).
\]
The expected utility of the plan is then
\[
E[P]=\sum_w P(w)\cdot u(w,P).
\]
The associated planning problem is to find the plan with the highest expected utility [1302.3597].

Poole also states a representation-size comparison. If \(n\) fluents each persist independently with some nontrivial probability under any action, the ICLsc representation requires only \(O(n)\) clauses, one stochastic frame axiom per fluent, whereas an equivalent probabilistic-STRIPS encoding must enumerate all \(2^n\) combinations of fluents that may be toggled by each action, producing an exponential blow-up. The same framework is described as related to structured POMDP representations, specifically two-slice temporal Bayes nets and action-network forms, but with greater representational economy because frame axioms need not be repeated for every action [1302.3597].

## 5. Dynamic stochastic utility and history-conditional random utility

Conditional random-utility representation also appears in dynamic choice, where conditioning is on histories rather than on parent variables or observables. In the dynamic stochastic utility model, time runs over \(t=1,\dots,n\), each period has a finite choice set \(X_t\), and menus are nonempty subsets \(A_t\in \mathcal M_t=2^{X_t}\setminus\{\varnothing\}\). A Dynamic Stochastic Choice Function is
\[
\rho=(\rho_1,\dots,\rho_n),
\]
with
\[
\rho_1:\mathcal M_1\to \Delta(X_1),
\]
and for \(t>1\),
\[
\rho_t(\cdot\mid h_{t-1}):\mathcal M_t\to\Delta(X_t),
\]
where the history \(h_{t-1}=(A_1,x_1;\dots;A_{t-1},x_{t-1})\) has positive probability. The induced joint choice probabilities are
\[
p(x_1,\dots,x_n;A_1,\dots,A_n)
=
\rho_1(x_1,A_1)\cdot \prod_{t=2}^n \rho_t(x_t,A_t\mid h_{t-1}).
\]
Letting \(P=\prod_{t=1}^n P_t\) be the product of strict-order spaces, stochastic utility is represented by a measure \(\mu\in \Delta(P)\) such that, for every \(t\) and every realized history,
\[
\rho_t(x_t,A_t\mid h_{t-1})
=
\mu(\{\succ:\text{ at history }h_{t-1}\text{ the top-ranked remaining element is }x_t\}\mid \{\succ:\text{ at }h_{t-1}\nu\text{ has survived}\}).
\]
The representation is therefore conditional on survival and on the observed choice-history itself [2102.00143].

The technical machinery for characterizing this representation uses Block–Marschak sums. The paper defines joint BM sums \(m(x,A)\), marginal BM sums \(m(x_{-n},A_{-n})\), and conditional BM sums \(m(x_n,A_n\mid (x,A)^{-n})\), together with an extrapolated conditional stochastic choice function. Möbius inversion yields
\[
p(x,A^C)=\sum_{D\le A} m(x,D),
\]
and a crucial decomposition identity is
\[
m(x_n,A_n\mid (x,A)^{-n})\cdot m(x_{-n},A_{-n})=m(x,A).
\]
The hierarchy “joint BM sums \(\Rightarrow\) marginal BM sums \(\Rightarrow\) conditional BM sums” provides the algebraic route from observed dynamic choice probabilities to a full conditional random-utility representation [2102.00143].

Several characterization theorems are reported. When \(|X_t|=3\) for all \(t\), a unique stochastic utility representation exists if and only if Joint Supermodularity and Marginal Consistency hold. When \(|X_t|=3\) for all \(t<n\) and the last-period set is arbitrary, stochastic utility is equivalent to Joint BM Nonnegativity plus Marginal Consistency; under strict positivity, the representing measure \(\mu\) has full support. An alternative four-axiom characterization replaces full joint conditions in the last period by partial marginal BM nonnegativity, partial conditional BM nonnegativity, \(({-}n)\)-Marginal Consistency, and \(({-}n)\)-Conditional Consistency. Without any cardinality restriction, stochastic utility is characterized by Joint Coherency in the De Finetti–Clark sense [2102.00143].

The concluding interpretation is explicit: one obtains a “full conditional random-utility representation” in which, at each history, the agent behaves as if a random continuation-utility draw selects the top element, and later periods reveal further coordinates of that same draw [2102.00143]. This places conditioning on histories at the center of dynamic random utility.

## 6. Neural random-utility representations

A contemporary extension of random-utility representation replaces closed-form latent utility functions with neural architectures while retaining the random utility maximization principle. In RUMnets, product features lie in \(\mathcal X\subset \mathbb R^{d_x}\), customer features in \(\mathcal Z\subset \mathbb R^{d_z}\), and classical random utility takes the form
\[
U(x,\epsilon,z,\nu)=\text{deterministic part}+\text{random noise},
\]
with an offered assortment \(A\subset \mathcal X\) and choice of the utility-maximizing alternative. RUMnets approximate the latent random fields \(\epsilon(\cdot)\), \(\nu(\cdot)\), and the utility \(U\) by feed-forward neural networks together with a sample-average approximation. For \(K\) independent samples of the two random fields and networks \(N_u\), \(N_{\epsilon,k_1}\), and \(N_{\nu,k_2}\), the sample-average utility is
\[
U_\theta^{(k_1,k_2)}(x,z)
=
N_u\bigl(x\oplus N_{\epsilon,k_1}(x)\oplus z\oplus N_{\nu,k_2}(z)\bigr).
\]
Adding i.i.d. white noise \(\delta_x\sim \mathrm{Gumbel}(0,1)/N^2\), the choice probabilities are
\[
p_j(x,z,A)
=
\frac{1}{K^2}\sum_{k_1,k_2}\mathrm{softmax}_j\bigl(\{U_\theta^{(k_1,k_2)}(x_i,z)\}_{i\in A}\bigr),
\]
where
\[
\mathrm{softmax}_j(u_i)_i=\frac{\exp(u_j)}{\sum_{i\in A}\exp(u_i)}.
\]
Equivalently, in the zero-noise limit they can be written as the probability that alternative \(x_j\) attains the maximum utility, averaged over the sampled latent fields [2207.12877].

Two formal claims define the significance of this construction. First, RUMnets sharply approximate the class of random utility maximization discrete-choice models: for any RUM choice model with continuous, bounded, uniformly continuous utility and i.i.d. random fields, and for every \(\eta>0\), there exist \(K,\ell,w,M\) and corresponding neural networks such that the induced RUMnet choice probabilities approximate the target model uniformly within \(\eta\). Second, every RUMnet is itself a random utility model: once the neural architecture is fixed and independent Gumbel shocks are added, the rule “choose \(\arg\max \tilde U(\cdot)\)” reproduces the RUMnet probabilities exactly [2207.12877].

The paper also reports a generalization bound for empirical-risk minimization under cross-entropy loss. For training data \(S=\{(y_t,z_t,A_t):t=1,\dots,T\}\), with probability at least \(1-\delta\),
\[
L^{true}(\hat\theta)
\le
L^{emp}(\hat\theta)
+
c_1\sqrt{\frac{\kappa^3\log d}{T}\,e^{2M}(2M)^\ell}
+
(8M+4\log\kappa)\sqrt{\frac{2\ln(4/\delta)}{T}},
\]
where \(\kappa=\max |A_t|\) and \(d\) is the input dimension [2207.12877]. Practical details given for this architecture include ReLU or ELU activations, typical depths \(\ell=3\ldots 10\), widths \(w=10\ldots 30\), mini-batch Adam on cross-entropy, label-smoothing, early stopping on a held-out validation set, optional weight decay or dropout, and implementations in Keras or PyTorch [2207.12877].

A plausible implication is that neural conditional random-utility representation preserves the economic interpretation of random utility while replacing analytically specified local structures with learned function classes. In that sense, it extends rather than abandons the core representational principle found in the earlier logical, graphical, and axiomatic models.

Source: https://www.emergentmind.com/topics/conditional-random-utility-representation