The Value of Information

https://www.frontiersin.org/files/Articles/805214/fenvs-10-805214-HTML-r1/image_m/fenvs-10-805214-g001.jpg

Imagine you’re about to buy a used car. It’s affordable, looks good, and you’re ready to make the purchase. But someone tells you that certain models of this car have a bug in their software, which can lead to engine failure. You don’t know whether this car has the faulty software or not, but you can pay a small fee to a dealership to check the version. Now you’re stuck wondering: should you pay for the check? Or skip it and risk buying a faulty car?

This is exactly the kind of decision that intelligent agents face every day. In the real world, AI rarely has access to perfect information. Much like humans, it operates under uncertainty. The question isn’t just what should I do now? it’s Should I also find out more before I decide?

This is the crux of the Value of Information. In this article, we’ll explore how AI can decide whether to gather more data before acting, how it measures the worth of that information, and what it means to act “intelligently” under uncertainty.

The Art of Asking the Right Questions

Let’s say you’re a doctor, and a patient comes in with a headache. You can’t know for sure what’s wrong just by looking. You could order an MRI, a blood test, or even ask about their sleep habits. But these tests take time, cost money, and sometimes even involve risks.

Just like a doctor, an AI agent must decide what to ask, when to ask, and how much it’s worth asking. The information itself can help it make better decisions, but that doesn’t mean it should always ask.

Not all information is equal. Sometimes, asking for more details doesn’t help at all. Other times, even a small piece of information can radically change the decision an agent would make. The challenge lies in knowing the difference.

Example: Should You Check the Car?

Let’s go back to the used car example. Suppose you’re paying $10,000 for the car. If the car has faulty software, and the engine breaks down, it’ll cost $4,000 to repair. There’s a 30% chance that this model has the faulty software, and if it does, there’s a 50% chance the fault will occur. You can get the software checked for $100. If you find that it does have the issue, you can negotiate the price down to $9,000.

You’re stuck with a question: do you test the software or not?

If you skip the test, your expected cost is: $10,000 + 0.3 × 0.5 × $4,000 = $10,600

But if you test, and then only buy if the price drops (or the car is safe), your expected cost might be lower.

This is how AI approaches the situation: it computes the expected cost or expected utility of both actions (testing or not testing) and picks the one with the better outcome.

• Expected cost if NOT checking: Price + P(faulty software) × P(fault will occur) × FixFault Cost
• Expected cost if checking: Price.(1 − P(faulty software)) + Pv × (Negotiated Price + P(fault will occur) × FixFault Cost) + Cost for Check

The key idea here is that information itself has value if it helps avoid worse outcomes.

What is the Value of Perfect Information (VPI)?

In the language of decision theory, we call this concept the Value of Perfect Information, or VPI. It’s a numerical measure of how much knowing something exactly would improve the agent’s decision-making.

Imagine an AI agent has already collected some evidence about the world. Based on what it knows, it computes the best possible action and the expected utility (or reward) it will receive from that action. This is called EU(α | e), where α is the best action given evidence e.

Now, imagine that the agent is magically told the exact value of an uncertain variable, say, whether or not the car has faulty software. With this new knowledge, the agent can recompute a possibly better action. It will also get a potentially higher expected utility, EU(α | e, Ej). Of course, it doesn’t know the exact value in reality, it can only imagine what would happen if it knew, and average over all possibilities.

So, the VPI is simply: VPI = Expected utility with extra information − Expected utility without it

We define,

• EU(α | e): Expected utility with current knowledge
• EU(α | e, Ej): Expected utility if we knew Ej
• VPIe(Ej) = Expected value of knowing Ej − value without it VPIe(Ej) = E[EU(α∣e,Ej)] − EU(α∣e)

If this number is greater than the cost of getting the information, then it’s worth asking. If not, it’s better to act with what you already know.

Learning What’s Worth Knowing

This framework turns an AI system into a thinking agent, not just a reactive one. It stops, evaluates its current understanding, and considers the future value of gaining more clarity. But it doesn’t just go around asking endless questions.

There’s a catch: gathering information isn’t always free. Whether it’s time, energy, money, or risk, every new observation has a price tag. AI doesn’t just ask for more, it asks wisely.

Simple rule:

If VPI ≥ Cost of Ej, then ask the question.

Otherwise, it’s not worth it.

Let’s build a basic AI agent that knows how to ask:

A Simple Strategy:

1. Let e = current knowledge
2. For all unknown variables Ej, calculate: VPIe (Ej) / Cost (Ej)
3. Pick the best one to ask.
4. If VPI > Cost, ask about Ej and update e.
5. If not, act based on current information.

There’s also a limit to how much it can do. Most AI systems use a “myopic” or short-sighted strategy: they only consider one variable at a time. They ask, “Is it worth learning this now?” without planning a long series of questions. It’s like a greedy algorithm, efficient, but not always perfect. Still, in many real-world systems, this works quite well.

Key Properties of VPI:

1. VPI is always ≥ 0 — learning never makes decisions worse.
2. VPI is order independent — doesn’t matter in which order you ask about variables.
3. But it’s not additive — asking about Ej and Ek separately is not the same as asking about both together.

You can think about it like this: having more information can never hurt decision-making. At worst, it doesn’t help. But it can’t make it worse.

Expert Systems and Real-World Applications

This idea of valuing information is at the heart of modern decision-theoretic expert systems. Think of a medical diagnosis system. It first builds a model of the world's diseases, symptoms, treatments, and outcomes. Then, it learns how these factors are related (probabilities), how severe or beneficial outcomes are (utilities), and finally, how to decide when and what to ask.

It doesn’t run every possible test on every patient. Instead, it calculates the expected value of each test and compares it with the cost, both financial and medical.

For instance, if a chest X-ray only helps in 5% of similar cases, and costs a lot, it might be skipped. On the other hand, if a simple blood test can significantly narrow down the possibilities, the system would recommend it.

These expert systems also undergo a process of refinement, comparing their decisions with those of human experts and continually improving over time. They perform a sensitivity analysis to ensure their recommendations don’t change drastically with small changes in the data. If they do, the model is unstable and needs better data or structure.

They use a decision-theoretic model with these steps:

1. Model the world: What are the causes (e.g., diseases), symptoms, and treatments?
2. Simplify the model: Work with experts to ignore or combine less important variables.
3. Assign probabilities: What’s the chance that disease A causes symptom X?
4. Assign utilities: How good or bad are outcomes (e.g., death = 0, recovery = 1)?
5. Verify & refine: Test the system’s output against expert diagnoses.
6. Do a sensitivity analysis: Slight changes in probability or utility shouldn’t lead to wild decisions. If they do, the model is unstable and needs work.

The Intelligence of Knowing When to Ask

The power of AI doesn’t just lie in calculating answers; it lies in knowing when it needs to know more. In uncertain environments, this becomes the true sign of intelligence.

By quantifying how useful information is (through the concept of VPI), AI systems can make thoughtful decisions. They don’t act blindly, nor do they seek information endlessly. They ask only when it’s worth it. They think before they ask. They act when they’re ready.

So the next time you see a chatbot asking for clarification, or your GPS asking whether you’d like a faster route, it’s not just being cautious. It might just be calculating the value of information, just like a good AI agent should.

• Smart agents don’t just act, they ask.
• But asking comes with a cost.
• So they calculate the Value of Information (VPI).
• If the VPI exceeds the cost, they go ahead and gather more info.
• If not, they act with what they have.

This makes AI strategic, not just reactive. Like a good doctor, it thinks before it probes.

In the next lesson, we’ll dive into Uncertainty and Probabilistic Reasoning, so stay tuned!

References:

• Russell & Norvig, Artificial Intelligence: A Modern Approach
• Lecture slides by Charith Chitraranjan