Influence Diagram

Why it matters

An influence diagram is a one-page map of a decision under uncertainty. It draws three kinds of thing — the choices you control (decisions), the chance variables you don’t (uncertainties), and the objective you’re trying to maximize (value) — and connects them with arrows that mean either “this influences that” or “this is known before that decision is made.” Its whole job is to show the structure of a hard decision compactly: what you decide, what’s uncertain, what you’ll learn before you commit, and how the outcome is finally scored — all without exploding into the thousands of branches a decision tree would need for the same problem.

For example: a company is deciding whether to build a new plant. The argument in the room keeps collapsing into “demand might be bad” versus “but the site is cheap.” Drawn as an influence diagram, the decision (which site, what scale) sits as a rectangle, demand and construction cost sit as ovals, and the payoff sits as a diamond — and an arrow appears that nobody had said aloud: a regulatory-approval oval feeds into the site decision, because approval risk depends on which site you pick. That single arrow reframes the choice. The question was never just cheap-versus-expensive; it was which site lets you wait for the demand signal before you commit the capital.

  • What it shows. The skeleton of a decision under uncertainty in one picture — the choices, the chance variables, what’s observed before deciding, and how the result is valued.
  • When to reach for it. A decision with real uncertainty and several moving variables, where you need to see the dependencies clearly before (or instead of) computing every branch.
  • How to read it. Read node shapes for kind — rectangle is a choice, oval is a chance variable, diamond is the objective — and arrows for relationship: into a decision means “known first,” into anything else means “affects it.”
  • What you’d miss without it. The information structure — which uncertainties resolve before which decisions. That timing is the most-often-omitted, most-decision-changing part of the problem, and the diagram forces you to draw it.
  • Where it misleads. A clean diagram can look like an answer when it is only a structure; the real numbers (the probabilities and the value function) live behind the nodes, and an arrow you forgot to draw is an assumption you forgot you made.

How to read it

Picture a small graph that flows left to right toward a single payoff on the right. Three shapes do the work. A rectangle is a decision — something you control, an alternative you get to pick (“which site,” “build now or wait”). An oval is an uncertainty — a chance variable you don’t control but that matters (“realized demand,” “construction cost,” “regulatory outcome”). A diamond (sometimes a hexagon or rounded box) is the value node — the single objective everything is ultimately scored against (“expected net present value”).

The arrows are where the grammar lives, and they mean different things depending on what they point at. An arrow into an uncertainty or the value node means influence: the source affects it — demand and cost both point into the value diamond because both move the payoff; site choice points into the regulatory oval because approval odds depend on the site. An arrow into a decision means something stricter — information: the source variable is known before that decision is made. So if a demand-signal oval has an arrow into a “scale” rectangle, it says you’ll observe demand before you choose scale. Drawing that one arrow is what turns a static gamble into a staged decision where waiting has value.

This is the high-level cousin of a decision tree. A tree makes the same decision fully explicit — every branch, every outcome, every payoff — which is great for computing an answer but sprawls fast: the same chance variable gets re-drawn under every branch, and the picture doubles with each new uncertainty. The influence diagram factors all that redundancy out. It shows the problem’s dependencies — what hinges on what — in a graph that stays roughly the size of the problem’s structure rather than the size of its enumeration, so it scales to decisions where a tree would be unreadable.

When to use it

The influence diagram belongs to Ora’s DECISION family — the diagrams built to frame and evaluate a choice — and within it, this is the structure-first member: the one you reach for to lay out a decision’s anatomy and dependencies compactly, especially when there are many variables and a tree would become unwieldy. Reach for it when the decision involves genuine uncertainty, when information will arrive before some decisions are made, when one decision feeds a later one, or when the outcome’s value is a function of several variables rather than a single obvious number.

Its closest relative is the decision tree, and the two are complementary views of the same problem rather than competitors. The influence diagram gives you the compact dependency map — good for framing, eliciting structure, and spotting missing arrows with a decision-maker. The decision tree gives you the explicit branches and payoffs — good for actually folding back the numbers and computing the optimal policy. A common workflow is to draw the influence diagram first to get the structure right, then unfold the relevant part into a tree (or solve the diagram directly) to compute. Skip the influence diagram when the decision is essentially deterministic (a flowchart of the process is enough), when the alternatives and trade-off are obvious and qualitative (a pro-con tree suffices), or when your aim is to model causation for explanation rather than to choose under uncertainty (that’s a causal DAG).

How Ora builds it

Ora produces an influence diagram from a semantic spec — a structured description that declares the typed nodes (the decision nodes you control, the uncertainty nodes you don’t, and the value node that scores the outcome) and a directed edge set whose meaning depends on the target: an influence arc into an uncertainty or value node (the source affects it) and an information arc into a decision node (the source is observed before that decision). That spec is rendered to a diagram with a graph-layout engine (Graphviz), flowing left to right and distinguishing the node types by shape, with an outline view and alt-text so the structure is legible without relying on shape and color alone.

The diagram is the visual face of Ora’s decision-architecture work: when you ask Ora to frame a hard choice under uncertainty, that mode names the decisions, the uncertainties, the information structure, and the value, and this artifact is how it shows the resulting skeleton. When the spec carries the quantities too — probability distributions on the uncertainties and a value function over its parents — the diagram becomes computable and can be solved for an optimal policy; when it doesn’t, it stands as a structural sketch for communication and elicitation.

The technique is the work of Ronald A. Howard and James E. Matheson at SRI and Stanford in the late 1970s and early 1980s. They needed a representation that scaled to industrial decisions where a decision tree would explode into thousands of branches, and their insight was that the tree’s redundancy could be factored into a compact graph if the semantics of the arrows were specified rigorously — informational into decisions, conditional into chance nodes, functional into value. Their canonical paper, circulated in 1981 and later reprinted in Decision Analysis (2005), defines exactly that grammar; the surrounding decision-analytic tradition is set out in Howard and Abbas’s Foundations of Decision Analysis.

  • Decision Tree — the explicit-branch member of the DECISION family: the same problem unfolded into every branch and payoff, the computational counterpart to this diagram’s compact structure.
  • Causal DAG — a directed graph that looks similar but answers a different question: it models causation for explanation and identification from data, not a choice under uncertainty.
  • Tornado Chart — the sensitivity companion: once an influence diagram is quantified, the tornado chart ranks which of its variables actually move the value.
  • Decision Architecture (mode) — the analytical operation that frames a decision’s choices, uncertainties, information, and value, and renders this diagram to show its structure.