Inference in Temporal Models: Filtering, Prediction, Smoothing, and Most Likely Explanation

In the previous article, we built the foundation for reasoning in a changing world: hidden states, evidence, transition models, sensor models, stationarity, and the Markov assumption.

Now comes the more practical question.

Once the model exists, what do we actually ask it to do?

In temporal probabilistic models, most useful questions fall into four inference tasks:

• Filtering: What is happening now?
• Prediction: What may happen in the future?
• Smoothing: What probably happened in the past?
• Most likely explanation: What full sequence of hidden states best explains the observations?

These tasks are the backbone of many real AI systems: robot localization, speech recognition, patient monitoring, weather tracking, aircraft tracking, and noisy signal reconstruction.

The mental model

A temporal model receives evidence over time.

At each step, the model tries to maintain a belief about the hidden state of the world.

For a hidden state X_t and observations e_{1:t}, the belief is usually a probability distribution, not a single answer.

The model is not just asking, “Is it raining?”

It is asking:

Given everything observed so far, how strongly should I believe each possible state?

That distinction matters. A belief state captures uncertainty.

The four major inference tasks

The same temporal model can answer different types of questions depending on the time point we care about.

Let us understand each one practically.

1. Filtering: estimating the current state

Filtering is the task of computing the current belief state.

Mathematically, filtering computes:

This means:

What is the probability distribution over the current hidden state, given all evidence observed so far?

In the umbrella world, filtering asks:

What is the probability that it is raining today, given all umbrella observations up to today?

Filtering is extremely practical because most intelligent systems need to know what is happening right now.

Examples:

• A robot estimates its current location.
• A doctor-support system estimates the current patient condition.
• A fraud detection system estimates whether the current transaction sequence is suspicious.
• A tracking system estimates the current position of an aircraft.

Filtering gives the system its current belief state.

2. Prediction: estimating a future state

Prediction asks about the future.

Mathematically, prediction computes:

This means:

Given all evidence up to now, what should we believe about a future hidden state?

In the umbrella world, prediction asks:

What is the probability that it will rain three days from now, given all umbrella observations so far?

Prediction is useful because decisions are usually made for the future.

Examples:

• A robot predicts where it will be after moving.
• A doctor predicts whether a patient may deteriorate.
• A weather system predicts tomorrow’s rain.
• A recommendation system predicts a user’s future interest.

Prediction is essentially filtering without new evidence.

We take the current belief and push it forward through the transition model.

3. Smoothing: estimating a past state using newer evidence

Smoothing is hindsight reasoning.

It computes a belief about a past state using evidence available up to the present.

Mathematically:

This means:

Given everything observed until now, what should we believe about a previous hidden state?

In the umbrella world, smoothing asks:

What is the probability that it rained last Wednesday, given all umbrella observations up to today?

This can be more accurate than the belief we had on Wednesday itself because we now have additional evidence from later days.

Practical examples:

• A medical system revises what the patient’s condition likely was yesterday after seeing today’s test results.
• A robot reconstructs where it probably was earlier after receiving later sensor readings.
• A speech recognition system uses later sounds to clarify an earlier ambiguous sound.

Smoothing is useful because later evidence can explain earlier uncertainty.

4. Most likely explanation: finding the best hidden sequence

Sometimes we do not only want the probability of one state.

We want the most likely sequence of hidden states that generated the observations.

This task is called most likely explanation.

Mathematically:

This means:

Among all possible hidden state sequences, which one best explains the observations?

In the umbrella world, suppose umbrellas are observed on the first three days and absent on the fourth day.

A likely explanation may be:

This task appears in many important AI systems.

In speech recognition, we observe a sequence of sounds and infer the most likely sentence. In communication systems, we receive a noisy signal and reconstruct the most likely original bit string.

This is not just state estimation. It is sequence reconstruction.

Filtering as recursive estimation

Filtering is especially important because it is the task we repeat as new evidence arrives.

The key idea is:

We should not recompute everything from the beginning every time new evidence arrives.

Instead, we maintain the current belief state and update it recursively.

If we already know:

and new evidence e_{t+1} arrives, we compute:

as a function of the previous belief and the new evidence:

This is called recursive estimation.

It has two steps:

1. Predict the next state from the current belief.
2. Update that prediction using the new evidence.

This predict-update cycle is one of the most important ideas in temporal AI.

Deriving the filtering update

The filtering target is:

First, split the evidence into old evidence and new evidence:

Using Bayes’ rule, we get:

Here, \alpha is a normalization constant. It ensures the final probabilities sum to 1.

Now apply the sensor Markov assumption.

The new evidence depends only on the current state X_{t+1}, not on the full previous evidence once X_{t+1} is known.

So:

Therefore:

The first term uses the new evidence. The second term is a one-step prediction.

Expanding the prediction step

Now we expand:

We condition on the previous hidden state X_t:

Using the Markov assumption:

So the prediction becomes:

Substituting this into the filtering update gives the final recursive filtering equation:

This equation has three important parts.

In plain English:

Why recursive filtering matters

Recursive filtering is efficient.

The model does not need to store every observation forever or recompute the full history at every step.

It only needs:

• the previous belief state,
• the transition model,
• the sensor model,
• the new evidence.

For discrete state variables, each update can be done in constant space with respect to time. That means the memory requirement does not grow just because the sequence gets longer.

This is why even a limited agent can keep tracking the current state over a long sequence.

Umbrella world: filtering step by step

Let us make the recursive update concrete using the umbrella world.

We represent the belief as a vector:

Assume the prior is:

This means the model initially assigns equal probability to rain and no rain.

The transition model is:

and the complementary transition structure gives:

The model also has persistence: rain tends to follow rain, and no-rain tends to follow no-rain.

The sensor model says umbrella observation is strong evidence for rain:

Day 1: prediction before seeing the umbrella

Before observing the day-1 umbrella evidence, we first predict rain on day 1 from the prior.

Using the prior [0.5, 0.5]:

This makes sense. Without evidence, we are still uncertain.

Day 1: update after observing umbrella

Now suppose we observe an umbrella on day 1:

We update using the sensor model:

Substitute the values:

Element-wise multiplication gives:

Normalize:

So after seeing an umbrella, the probability of rain becomes much higher.

The evidence changed the belief.

Day 2: predict from the updated day-1 belief

Now we move to day 2.

Before seeing the day-2 umbrella observation, we predict using the transition model:

Using the day-1 belief [0.818, 0.182]:

Even before observing day-2 evidence, rain is still more likely than the original prior. Why?

Because rain tends to persist in this model.

Day 2: update after observing umbrella again

Now suppose an umbrella is observed again on day 2:

Update again:

Substitute the values:

Element-wise multiplication gives:

Normalize:

Now the probability of rain is even higher than it was after day 1.

Two umbrella observations provide stronger evidence than one.

Visualizing the umbrella filtering flow

The computation can be understood as messages moving along a chain.

The pattern is always the same:

That pattern is the heart of filtering.

Prediction as filtering without new evidence

Prediction uses the first half of the filtering update.

Filtering does two things:

1. Predict forward.
2. Update using new evidence.

Prediction only does step 1.

Mathematically:

This answers:

Given what we believe now, what should we expect next?

If we predict multiple steps into the future without receiving new evidence, the distribution may gradually converge to a fixed point.

This fixed point is called the stationary distribution of the Markov process.

The rough time taken to approach this fixed point is called the mixing time.

In practical terms:

The further into the future we predict without new evidence, the more the model forgets the specific observations and settles into its long-term behavior.

A practical way to remember all four tasks

Here is the simplest way to remember the difference.

How this connects to modern AI systems

These ideas may look classical, but they map beautifully to modern AI systems.

AI agents

An AI agent has to track the current task state, user goal, tool results, pending approvals, failed attempts, and environment changes.

Filtering helps answer:

What is the current state of the task?

Prediction helps answer:

What will likely happen if the agent takes this action?

Smoothing helps answer:

Given what happened later, where did the workflow probably go wrong?

Most likely explanation helps answer:

What sequence of internal states best explains this conversation or tool trace?

Robotics

A robot uses filtering to estimate where it is now, prediction to estimate where it will be after movement, and smoothing to reconstruct previous positions after receiving later sensor data.

Healthcare AI

A patient-monitoring system uses filtering for current condition estimation, prediction for future risk, and smoothing to revise earlier assumptions after new test results arrive.

Speech recognition

A speech system observes sounds and infers the most likely word or sentence sequence. That is a most-likely-explanation problem.

Shareable takeaway for LinkedIn and Medium

If you are learning or teaching Advanced AI, remember this simple frame:

This series is designed to make classical AI concepts practical for modern AI engineering: agents, robotics, healthcare AI, tracking systems, speech recognition, and decision-making under uncertainty.

If this helped you, share the article with one learner or engineer who is trying to build stronger AI fundamentals.

Key takeaways

• Temporal inference starts after we define hidden states, evidence, transition models, and sensor models.
• Filtering computes the current belief state: P(X_t \mid e_{1:t}).
• Prediction computes future belief: P(X_{t+k} \mid e_{1:t}).
• Smoothing revises a past belief using later evidence: P(X_k \mid e_{1:t}).
• Most likely explanation finds the best hidden state sequence: \arg\max_{x_{1:t}} P(X_{1:t} \mid e_{1:t}).
• Filtering is recursive: it updates the belief state as new evidence arrives.
• The filtering update has two steps: prediction and evidence update.
• The transition model moves belief forward in time.
• The sensor model adjusts belief using new observations.
• Long-term prediction may converge to a stationary distribution.

Final thought

Temporal inference is not about memorizing formulas. It is about learning how to update belief as the world changes.

A good AI system does not simply react to the latest observation. It combines:

• what it believed before,
• how the world tends to evolve,
• and what the new evidence suggests.

That is why filtering, prediction, smoothing, and most likely explanation are foundational tools for building AI systems that operate in the real world.