KomuraSoft LLC
Chapter 2

Predict step — predicting the next state from the model

Under the steady-state model the value does not change, but uncertainty grows over time. Verify P⁻ = P + Q with a case study.

The Kalman filter does not use each observation as is. It first uses a model to predict "this is what it should be next," and only then mixes in the observation.

This course uses the minimal model

Here we only use a one-dimensional steady-state model. The steady-state model is a model that assumes the target does not move (its position does not change), so we set "the next time step will, for now, be the same value as the previous one."

x̂⁻ = x̂

Thanks to this simplification we can focus on what really matters right now: the division of labor between prediction and update.

Uncertainty grows over time

Even if the value itself does not change, uncertainty grows the further into the future we go. We express this with the variance P and update it in the prediction step with the following formula.

P⁻ = P + Q

Q expresses "how easily the model tends to be off." The longer we go without new observations, the more P grows.

Check 1 — Trace the prediction step by hand

From the previous-step estimate, build the predicted value x̂⁻ and predicted variance P⁻.

Q1. With the previous estimate x̂ = 12, P = 4, and Q = 1.5, what is the predicted value x̂⁻ under the steady-state model?

Q2. Under the same conditions, what is the predicted variance P⁻?

Q3. With initial variance P₀ = 2 and Q = 0.5, if we only predict three times in a row without any observations, what is the variance?

The size of Q changes the behavior

Small Q case
The model is assumed to be fairly accurate — for example, a camera fixed on a tripod where the target barely moves. It becomes easier to keep trusting the prediction.
Large Q case
The model is assumed to be unreliable — for example, tracking a target that suddenly changes direction. When a new observation arrives, the estimate is pulled strongly toward the observation.

Check 2 — Grasp the role of Q

Compare Case A and Case B with different values of Q, then put the meaning of Q back into words.

Q1. With previous variance P = 1 and Q = 0.01, what is the predicted variance P⁻?

Q2. With the same P = 1 and Q = 4.0, what is the predicted variance P⁻?

Q3. Which statement best captures a case where Q is very small?

Takeaway of this chapter

Prediction is the operation "the value does not change, but uncertainty grows a little." In the next chapter we look at how much to move that prediction once an observation arrives.