🧭 Maybe What's Missing When You "Can't Decide" Is Just One of Five Elements

Decision Making / Data Science / AI / Statistics

— The Decision Map as a Scaffold for Thinking —

Introduction

You try to decide something, and you go round in circles.

“What should I have for lunch today?” “Which method should I use for this research?” “How should I delegate this to AI?”

The scale is completely different each time, yet the texture of the struggle feels similar. All sorts of information is floating around in your head, but the final call won’t come. Maybe there’s a single reason common to all of these states.

This blog post is about a tool for making that “common reason” visible — a way of thinking called the decision map. It’s a concept put forward by Professor Toshiyasu Matsushima of Waseda University, in the context of organizing data science as “a discipline dealing with the clear, rational process of making decisions from data,” and it’s laid out systematically in his papers and book series.

It got long, so I split it into four parts.

• Part 1: What is a decision map? — The five elements
• Part 2: Statistics and optimization can both be written as decision maps
• Part 3: There is no decision-making without subjectivity — what does a scientific attitude mean?
• Part 4: The skill you need in the AI era — standing on the side that designs the five elements

Read them in order, or just read whichever part interests you.

Part 1: What Is a Decision Map? — The Five Elements

Most of “struggling to decide” is just one of the five elements not yet put into words

A decision map is the idea of treating every decision as a “map” (function) that takes an input and picks one item out of an output set.

That map is described using the following five elements.

If even one of the five elements is missing, the decision can’t be completed. Put the other way around: when you feel you “can’t decide,” it’s usually because one of the five elements hasn’t been put into words yet.

Example: deciding what to have for lunch

Even a trivial choice has all five elements once you write it out.

And you notice that the usual cause of “hesitating” is that the weighting of the evaluation criteria hasn’t been fixed. When “won’t make me sleepy” and “tasty” collide, you can’t decide because you haven’t fixed which one takes priority. This isn’t a problem with the goal or a lack of input — it’s an evaluation-criteria problem.

Write it out, and you see what the struggle really is.

Career worries have the same structure

Career worries are, more often than not, a problem where the goal itself hasn’t been fixed. Even the mismatch over whether to prioritize “growth opportunity” or “salary” as an evaluation criterion is rooted in the goal layer.

Do you want to “grow,” or do you want “stability”? If this is wavering, the evaluation criteria will waver forever, and no answer will come no matter how long you stare at the options.

Once you write the five elements out on a single page, you can see which element you usually haven’t put into words. That is the real reason your decision-making has stalled.

That’s Part 1. Next comes the reach of this way of seeing things.

Part 2: Statistics and Optimization Can Both Be Written as Decision Maps

What looks like “a method” is, in fact, always a decision map

This is where the decision map really shows its worth.

In fact, what get called “methods” — regression analysis, optimization, machine learning — can all be described as decision maps. This is the starting point of Professor Matsushima’s paper, and it’s the core of a viewpoint that captures data science in a unified way.

Example: multiple regression analysis

Given you want to predict a target variable $y$ from explanatory variables $x$, you seek a linear function $y = a^\top x$. Written in five elements:

What’s beautiful here is the coupling between the setting and the evaluation criteria. Because the setting fixed “errors are homoscedastic and normally distributed,” maximizing likelihood turns out equivalent to minimizing squared error. Change the setting to a Laplace distribution, and under the same likelihood maximization you get minimization of absolute error instead.

The five elements aren’t independent. They’re coupled to one another. Good design sense comes down to whether you can stay aware of that coupling.

Ridge regression, Lasso, AIC — all written in the same framework

In Professor Matsushima’s paper, ridge regression, Lasso, AIC, variable-selection problems, and even variational Bayes methods are all organized uniformly as pairs of a decision map and an evaluation function. You come to see things not as “memorize a pile of methods” but as “what family of methods appears as you vary the evaluation function or the setting of the same map.”

In the book series Introduction to Data Science, this same view lays out descriptive statistics, inferential statistics, and Bayesian statistics — fields that take different stances — side by side, in the common language of the decision map.

Even “the same map” changes meaning when the goal above it differs

This is especially interesting.

Professor Matsushima points out that the very same decision map and evaluation function for multiple regression admits two different stances.

• Stance 1: I want to describe the data I obtained well (the descriptive-statistics stance)
• Stance 2: I want to know about the population/probability distribution behind the data (the mathematical-statistics stance)

Written as an equation, the two look identical, yet the higher-level goal — what you’re deciding for — is different. Under Stance 2, the output $a$ isn’t just a coefficient; it’s interpreted as an estimator of a population parameter.

In other words, a decision map has a hierarchy above it. Above the “goal” there is a “higher-level goal,” and above the evaluation criteria there is the question of grounds: “why is this the right evaluation?”

Hierarchical structure — fractal nature

A decision map is nested (fractal).

Top level:   Want the company to survive 10 years from now
   ↓ toward that
Strategy:    Want to strengthen the engineering organization
   ↓ toward that
Tactics:     Want to hire 3 senior engineers this term
   ↓ toward that
Implementation: Should we hire Person A?

Each layer is its own decision map with its own five elements. The lower layer’s “goal” is nothing more than “one item in the output set” from the upper layer’s point of view.

If you agonize only over the implementation and lose sight of the top level, your judgment wobbles. If you only talk about the top level and never decide on the implementation, you can’t move forward. Without being aware of which layer you’re currently deciding in, the conversation won’t line up.

And, “how to design the five elements” is itself, once again, a decision map (a meta-decision). If you feel you “can’t quite decide,” the better move isn’t to add more options one layer down — it’s to doubt the map one layer up: “is the evaluation criterion even right to begin with?”

That’s Part 2. A mountain of methods, and the common language of the decision map changes the whole view.

Next is the topic of subjectivity and objectivity. This is probably where the decision map does its best work.

Part 3: There Is No Decision-Making Without Subjectivity — What Does a Scientific Attitude Mean?

“We decided objectively” is, in almost every case, not a lie but imprecise

This might be the decision map’s most practical contribution.

Professor Matsushima’s paper makes an extremely important point about the limits of data science. To summarize:

The mathematical description of the decision map and the evaluation function is clear and rational, but the higher-level notions — “why this goal,” “why is this the right evaluation,” “what assumptions are we allowed to make” — must be decided from outside the system of data science. They depend on the intent of the deciding party and on domain expertise in the field concerned.

This isn’t a criticism that “data science is vague.” It’s a statement of the precondition for using data science correctly.

Where in the five elements is subjective, and where is objective

Color-coding the five elements of the decision map by subjective/objective gives you this:

Data and facts bear directly on only the “input” — the other four elements are, fundamentally, designed by human subjectivity.

The phrase “we decided based on the data” is, in almost every case, not a lie, but it’s imprecise. In reality, you’re just processing data inside a frame that was assembled subjectively.

Example: what gets called “data-driven hiring”

Only 2 of these 6 items are based on data. All the rest is a frame someone designed subjectively.

What a scientific decision is — not eliminating subjectivity, but making the boundary explicit

This is the core point.

Not eliminating subjectivity, but making explicit the boundary between the subjective and the objective — that is the scientific attitude toward decision-making.

Once you write out the five elements:

1. You can see where the subjective parts are (goal, setting, evaluation criteria, output set)
2. You can see where the objective parts are (input data, the computation process)
3. You can own the subjective parts as “this is my judgment”
4. You can present the objective parts as “this part is verifiable”

Someone who insists “we decided objectively” is far less trustworthy than someone who separates things out and says “up to here is subjective, from here on is data.” Whether you run things aware of this or unaware of it makes a completely different difference to the quality of your judgments and to how improvable they are.

The decision map is a device that stops “data-driven” from being used as a decorative phrase, and forces you to ask again how much of it is really being driven [by data].

That’s Part 3. The last part is about the AI era.

Part 4: The Skill You Need in the AI Era — Standing on the Side That Designs the Five Elements

AI can only handle the “contents” of a decision

If you’ve read this far, you may already see this conclusion coming.

AI (machine learning, LLMs, optimization solvers) can, in essence, only do the following:

Given the “goal, setting, evaluation criteria, output set” among the five elements, take an input and choose an output.

In other words, AI cannot design the five elements. Delegating a decision to AI is equivalent to handing it the five elements.

This is a logical consequence of the decision map. No matter how smart AI gets, designing “this is what the goal should be” or “this is the evaluation criterion” ultimately stays on the human side. The responsibility for designing that cannot, in principle, be dumped onto AI.

The gap between people who can hand over the five elements and people who can’t

This isn’t the surface-level skill of “writing a good prompt.” It’s the ability to treat decision-making itself as a structure.

One more step — designing AI agents

Everything so far is about the stage where a human calls on AI for individual decisions.

One step further, in the AI agent — a system that autonomously breaks down tasks, calls tools, and chooses and executes actions — the structure of the five elements gets lifted up one level.

The agent assembles the five elements of each individual task itself, given a goal. But what makes the agent possible in the first place is the “five elements of the agent,” designed by a human.

Agent development has entered a stage where it’s not “handing an individual decision over to AI,” but “designing the deciding subject itself.”

The work left for humans is clear.

• Designing the agent’s five elements (especially the goal, the evaluation criteria, and action constraints)
• Structurally monitoring whether the individual-task five elements the agent assembles are running out of control
• When the agent fails, isolating which element, at which layer of the five elements, went wrong

Someone who holds a decision map can, in the agent era, stand on the side that designs the agent. Someone who doesn’t just ends up watching the agent’s output. This gap is about to become decisively large.

Conclusion — Replacing “Struggling” with “Designing”

This got long. Let me compress the key points onto one page at the end.

1. Every decision can be written using the five elements: goal, setting, evaluation criteria, input, output set
2. Data science, statistics, optimization, and machine learning are all just special cases of this
3. The five elements are hierarchically nested (fractal)
4. There is always a meta-decision that determines the five elements themselves
5. Data and facts bear directly on only the “input.” The other four elements are human subjectivity
6. A scientific decision is not the elimination of subjectivity, but making the boundary between subjective and objective explicit
7. In the AI era, the people who create value are those who can hand the five elements over to AI — and further, those who can stand on the side that designs the five elements of the AI agent itself

If you’re going to try just one thing starting tomorrow, try this:

Pick one decision in front of you, and write the five elements out on a single page.

If there’s an element that’s hard to write, that’s the part you usually avoid putting into words. That element is the real reason your decision-making has stalled.

Having a decision map means you can replace “struggling” with “designing.” That alone raises both the speed and the quality of your decisions.

References

• Toshiyasu Matsushima, “How Should We Grasp and Teach Data Science? — Drawing on Waseda University’s Efforts,” Operations Research, Vol. 65, No. 10, 2020, pp. 544–550. https://orsj.org/wp-content/corsj/or65-10/or65_10_544.pdf 　— The paper that clearly situates the concept of the decision map within the system of data science. The core source cited in this article.

• Waseda University Data Science Education Team (author), supervised by Toshiyasu Matsushima, Introduction to Data Science I — Foundations of Data-Based Decision Making (Data Science Library 1), Science-sha, 2022. ISBN: 978-4-7819-1540-1. 　— A textbook that systematically explains descriptive statistics, inferential statistics, and Bayesian statistics within the unified framework of the decision map. Recommended if you want to properly study the decision map — it’s systematic. The series currently runs to Volume IV.