Decision Analysis: Multi-Attribute Utility and Decision Networks

The utility-theory lesson built one agent that maximizes expected utility, alone, against a non-strategic environment. It left the practical machinery of decision analysis unbuilt: how to specify utility when an outcome has many attributes, how to lay a decision problem out as a network you can evaluate, and how to price the information an agent might buy before deciding. This lesson supplies that machinery and sets up the strategic, multi-agent decisions of the next lesson.

Utility over many attributes

If an outcome is described by several attributes, a table holding a preference for every combination of their values grows exponentially. The remedy is structure in the agent's preferences — regularities that collapse the big table into a few small ones, exactly as conditional independence collapsed a joint distribution earlier in the module. This section works from the crudest tool (dominance, which needs no numbers) up to the compact additive form.

Real decisions rarely turn on one number. Siting an airport weighs construction disruption, land cost, distance from population, noise, and safety all at once. A public-policy choice weighs money against lives. Multiattribute utility theory handles outcomes described by a vector of attributes $\mathbf{X}=⟨X_1,\dots ,X_n⟩$, each $x_i$ a number or an ordered discrete value.¹ We assume utility increases monotonically in each attribute (more absence-of-noise is better; if an attribute like temperature has an interior peak, split it into two monotone attributes measuring distance from the ideal on each side).

Dominance

Sometimes you can decide without combining attributes into a single utility. Suppose airport site $S_1$ costs less, makes less noise, and is safer than $S_2$. Then $S_1$ strictly dominates $S_2$: it is at least as good on every attribute, so $S_2$ can be dropped from consideration no matter how the attributes trade off. Strict dominance rarely picks a unique winner, but it prunes the field cheaply.

That is the deterministic case, where attribute values are known exactly. Under uncertainty each option induces a distribution over each attribute, and a direct analog holds: $S_1$ strictly dominates $S_2$ if every possible outcome of $S_1$ beats every possible outcome of $S_2$. This almost never happens. The useful generalization is stochastic dominance. Suppose cost at $S_1$ is uniform on $[\text{\$}2.8\text{B},\text{\$}4.8\text{B}]$ and cost at $S_2$ is uniform on $[\text{\$}3\text{B},\text{\$}5.2\text{B}]$. Plotting cost as a negative value, the two distributions overlap, yet $S_1$'s is shifted toward lower cost throughout, and any agent whose utility decreases with cost prefers $S_1$ — without ever pricing money.

Formally, if actions $A_1,A_2$ induce densities $p_1,p_2$ on attribute $X$, then $A_1$ stochastically dominates $A_2$ when its cumulative distribution never exceeds the other's,

and the consequence is a theorem: if $A_1$ stochastically dominates $A_2$, then for every monotonically nondecreasing utility function $U(x)$, the expected utility of $A_1$ is at least that of $A_2$. So a dominated action can be discarded without knowing $U$ at all. The condition can often be settled qualitatively — if transport cost rises with distance to the supplier and $S_1$ is closer, $S_1$ dominates on cost — and there are algorithms that propagate such qualitative facts through qualitative probabilistic networks to reach rational decisions with no numeric values whatsoever.

Preference structure and the additive form

When dominance leaves several contenders you must combine attributes. To tabulate $U(x_1,\dots ,x_n)$ directly over $n$ attributes with $d$ values each takes $d^n$ numbers — the worst case, an agent whose preferences have no regularity at all. Multiattribute utility theory rests on the supposition that real agents have far more structure than that, and the tool for exploiting it is the representation theorem: identify a regularity in the agent's preferences and derive a compact utility from it. Formally we hope to show that an agent with a certain preference structure has a utility function

where $F$ is some simple combining function — ideally addition. The parallel to a Bayesian network is exact: there, a structural regularity (conditional independence) let us decompose a joint probability into small local terms; here, a structural regularity in preferences will decompose a utility into small local terms.

Preferential independence

Begin with the deterministic case, where the agent has a value function$V(x_1,\dots ,x_n)$ over outcomes known for sure. The basic regularity is preferential independence.

Take the airport attributes Noise, Cost, Deaths. Suppose you prefer a site with $20,000$ people in the flight path at a $\text{\$}4$ billion construction cost over one with $70,000$ people at $\text{\$}3.7$ billion when safety is $0.06$ deaths per million passenger-miles in both cases. If you hold the same preference when safety is instead $0.12$, and again when it is $0.03$ — and likewise for every other pair of Noise/Cost values — then $\{Noise,Cost\}$ is preferentially independent of Deaths. The trade-off between noise and cost is fixed; only the safety level riding alongside changes, and it does not move the trade-off.

Symmetrically, $\{Cost,Deaths\}$ is preferentially independent of Noise, and $\{Noise,Deaths\}$ of Cost. When every subset of attributes is preferentially independent of the rest, the set exhibits mutual preferential independence (MPI). MPI says that each attribute may matter to the overall value, yet none of them changes how you trade off the others against each other.

The additive-value theorem

MPI is a strong assumption, but it yields a simple form, due to a theorem of the economist Gérard Debreu.

For the airport this might be

Assessing an additive value function means eliciting $n$ one-dimensional functions $V_i$ instead of one $n$-dimensional table — typically an exponential reduction in the number of preference questions the analyst must ask. And the additive form is robust: even when MPI does not hold exactly at extreme attribute values, it often approximates the true preferences well, provided the violations occur in ranges of the attributes that seldom arise in practice.

MPI can genuinely fail, and the failure is instructive. At a medieval market you are buying hunting dogs, chickens, and wicker cages. The dogs are valuable — but with too few cages, the dogs eat the chickens, so the trade-off between dogs and chickens depends strongly on the number of cages. That three-way interaction breaks preferential independence and, with it, additivity: you cannot score dogs, chickens, and cages separately and sum, because the value of dogs is entangled with how many cages you hold.

Utility independence under uncertainty

Everything above concerned value functions over sure outcomes. When outcomes are uncertain the agent chooses among lotteries, and we need the analogous structure on preferences over lotteries. Utility independence extends preferential independence to cover them.

MUI is the uncertain-world counterpart of MPI, and it yields its own representation theorem. Where MPI gave a purely additive value function, MUI gives a multiplicative utility function. For three attributes, writing $U_i$ for $U_i(x_i)$,

The expression looks busy, but it contains only three single-attribute utilities $U_1,U_2,U_3$ and three constants $k_1,k_2,k_3$. In general an $n$-attribute MUI problem needs $n$ single-attribute utilities and $n$ constants, each of the utility functions developed independently of the others, and the combination is guaranteed to reproduce the correct overall preferences. What the additive form lacks is the cross-terms $U_i{U}_j$ — they capture how the attributes interact under risk — and with additional independence assumptions those terms vanish and the utility collapses to the purely additive $U={\sum }_i{k}_i{U}_i$.

Decision networks

A decision network — the more descriptive name for an influence diagram — lays a decision problem out as a graph, extending a Bayesian network with two new node types.² It captures the agent's current state, its possible actions, the state its action produces, and the utility of that state, all in one picture you can evaluate mechanically.

Three node types appear:

• Chance nodes (ovals) are random variables, exactly as in a Bayesian network, each carrying a conditional distribution indexed by its parents. Parents may be chance or decision nodes. Air Traffic, Litigation, Construction describe the current state; Deaths, Noise, Cost describe the outcome and depend on the chosen site.
• Decision nodes (rectangles) are the points where the agent chooses. Here Airport Site takes a different value for each candidate site, and that choice influences cost, safety, and noise. We treat the single-decision case; sequential decisions are the province of MDPs.
• Utility nodes (diamonds, also called value nodes) hold the utility as a function of the parents that directly affect it — a table, or a parameterized additive/linear function of the attribute values.

A common shorthand omits the outcome-state chance nodes and wires the utility node straight to the decision and current-state nodes. The utility node then encodes the expected utility of each action given the state — an action-utility function, the same object called a Q-function in reinforcement learning. Because outcome nodes like Noise and Cost refer to future states and can never be observed as evidence, this compiled form is always usable when the full one is; it has fewer nodes but hides the outcome structure, so a change in aircraft noise now has to be edited into the action-utility table rather than a single conditional distribution.

Evaluating a decision network

Once the decision node is fixed to a value it behaves like an observed chance node, and inference in the resulting Bayesian network gives the expected utility of that action. Trying each action and keeping the best gives the algorithm.

This is a direct extension of Bayesian-network inference, and it slots into a utility-based agent unchanged. What makes the sequential problem harder — several decisions in a row, each observing the outcome of the last — is the province of MDPs; a single decision node needs only one inference sweep per action.

The value of information

The evaluation above assumed the agent already holds all available evidence. In practice one of the most important parts of deciding is knowing what questions to ask. A doctor cannot run every test at once; tests cost money and time and sometimes carry risk, so their worth depends on whether the result would change the treatment plan and by how much. Information value theory lets an agent decide which observations to acquire, treating the observation of a chance variable as a simplified action that changes only the agent's belief state, not the world.

A worked example

An oil company can buy one of $n$ indistinguishable ocean-drilling blocks. Exactly one block holds oil worth $C$ dollars; the rest are worthless. Each block sells for $C/n$, so a risk-neutral company is indifferent between buying and not buying — its expected profit is zero either way.

Now a seismologist offers a survey of block $3$ that reveals definitively whether it holds oil. What is that survey worth? Reason through what the company would do with the answer:

• With probability $1/n$ the survey says block 3 has oil. The company buys it for $C/n$ and profits $C-C/n=(n-1)C/n$.
• With probability $(n-1)/n$ the survey says block 3 is dry. The company now buys one of the other blocks, whose chance of oil has risen from $1/n$ to $1/(n-1)$, for an expected profit of $C/(n-1)-C/n=C/(n(n-1))$.

Averaging over the two survey outcomes,

The information is worth exactly $C/n$ — the price of a block, and the entire expected profit the company can now make. Its value comes from letting the company tailor its action to the actual situation instead of doing what is best on average across situations.

The VPI formula

Let the agent's current evidence be $\mathbf{e}$. The expected utility of the best action now is

and after learning $E_j=e_j$ the best action might change, giving $EU({\alpha }_{e_j}\mid \mathbf{e},e_j)$. But $E_j$ is a random variable whose value is currently unknown, so we average over its possible values $e_{jk}$ using current beliefs. The value of perfect information is

In words: the expected utility of acting after the observation, minus the expected utility of acting now. Three regimes illustrate it. Choosing between a safe highway ($a_1$) and a winding dirt road ($a_2$) over a mountain range, the value of a satellite report on road conditions depends on how the resulting utilities $U_1^{\prime },U_2^{\prime }$ are distributed.

In short: information has value to the extent that it is likely to cause a change of plan, and to the extent that the new plan is significantly better than the old one. In case (a) no report changes the decision; in (b) it might, and the stakes are high; in (c) it might, but the stakes are tiny.

Properties, and an agent that gathers information

VPI has two general properties.

It is a statement about expected value, not actual value — a misleading test can lead to a worse plan (a false positive prompting needless surgery), but that does not mean the test should be skipped. The second property is that VPI is not additive — learning $E_j$ can raise or lower the value of later learning $E_k$ — but it is order independent:

Order independence is what distinguishes sensing actions from ordinary actions and simplifies planning a sequence of observations. An information-gathering agent uses VPI as its guide: request the observation with the best value-per-unit-cost, until no observation is worth its cost.

Because it evaluates VPI as though only one variable will ever be observed, this design is myopic — the same shortsighted heuristic as greedy search. It can undervalue a sequence of observations that only jointly pay off, but in practice it works well; myopic information-gathering has been shown to outperform expert physicians at selecting diagnostic tests.

That completes single-agent decision analysis. Every tool so far assumed a non-strategic environment. When outcomes depend on other rational agents, whose choices react to the agent's own, a different theory is needed.

This continues in Game Theory and Mechanism Design, which studies decisions among agents (Nash and maximin equilibria) and then inverts the question to engineer rules under which selfish play yields a good collective outcome.