Weighted Averages¶
Level: Middle School (Ages 13-15)
The Problem with Regular Averages¶
Imagine you're trying to find out the temperature outside. You have: - Your guess: 80°F (you're just guessing) - A cheap thermometer: 72°F (somewhat reliable) - A weather station: 70°F (very reliable)
Regular average: (80 + 72 + 70) ÷ 3 = 74°F
But wait! Should your wild guess count as much as the weather station? Probably not!
What is a Weighted Average?¶
A weighted average gives more importance to some numbers than others.
Formula¶
Example: Temperature with Weights¶
Let's give weights based on reliability: - Your guess: 80°F, weight = 1 (not very reliable) - Cheap thermometer: 72°F, weight = 3 (somewhat reliable) - Weather station: 70°F, weight = 6 (very reliable)
Notice: 71.6°F is much closer to the weather station (70°F) than to your guess (80°F)!
Why Weights Matter¶
Example 1: School Grades¶
Your teacher calculates your grade: - Homework: 85% (weight = 20%) - Quizzes: 90% (weight = 30%) - Final exam: 78% (weight = 50%)
Regular average: (85 + 90 + 78) ÷ 3 = 84.3%
Weighted average:
The final exam counts more, so it pulls your grade down!
Example 2: Combining Measurements¶
You measure the length of a table: - Measurement 1: 48.2 inches (uncertainty ± 0.5 inches) - Measurement 2: 48.0 inches (uncertainty ± 0.1 inches)
Which measurement should you trust more? Measurement 2! It has lower uncertainty.
These sensor-fusion formulas assume unbiased observations of the same quantity with independent errors and known variances. Here ± denotes one standard deviation. Correlated measurements need their joint covariance; an arbitrary confidence score is not a variance.
Weights from Uncertainty¶
Here's the key insight: Lower uncertainty = Higher weight
The Rule¶
Or in math terms:
Example: Table Measurements¶
Measurement 1: 48.2 inches, uncertainty = 0.5 - Variance = 0.5² = 0.25 - Weight = 1 ÷ 0.25 = 4
Measurement 2: 48.0 inches, uncertainty = 0.1 - Variance = 0.1² = 0.01 - Weight = 1 ÷ 0.01 = 100
Measurement 2 gets 25 times more weight! (100 vs 4)
Weighted average:
The result is very close to the more accurate measurement!
The Kalman Filter Way¶
The Kalman filter uses weighted averages to combine: 1. Predictions (what you expect) 2. Measurements (what you observe)
Each has its own uncertainty, so each gets its own weight!
Example: Tracking a Car¶
Prediction: The car is at position 100 meters (uncertainty ± 5 meters) Measurement: GPS says 95 meters (uncertainty ± 10 meters)
Which should you trust more? The prediction! It has lower uncertainty.
Prediction weight: 1 ÷ 5² = 1 ÷ 25 = 0.04 Measurement weight: 1 ÷ 10² = 1 ÷ 100 = 0.01
Weighted estimate:
The estimate is closer to the prediction (100m) than the measurement (95m) because the prediction is more certain!
Kalman Gain: The Magic Number¶
The Kalman filter uses something called Kalman Gain (K) to decide how much to trust the measurement vs. the prediction.
Formula¶
What K Means¶
- K = 0: Don't trust the measurement at all (use prediction)
- K = 1: Don't trust the prediction at all (use measurement)
- K = 0.5: Trust both equally
Example: Car Tracking¶
Prediction uncertainty: 5 meters Measurement uncertainty: 10 meters
Update formula:
New Estimate = Prediction + K × (Measurement - Prediction)
= 100 + 0.2 × (95 - 100)
= 100 + 0.2 × (-5)
= 100 - 1
= 99 meters
Same answer as the weighted average!
Understanding Kalman Gain¶
Case 1: Very Certain Prediction¶
Prediction: 100 ± 1 meter Measurement: 95 ± 10 meters
K is very small! Trust the prediction more.
Very close to the prediction (100m)!
Case 2: Very Certain Measurement¶
Prediction: 100 ± 10 meters Measurement: 95 ± 1 meter
K is almost 1! Trust the measurement more.
Very close to the measurement (95m)!
Case 3: Equal Uncertainty¶
Prediction: 100 ± 5 meters Measurement: 95 ± 5 meters
K is 0.5! Trust both equally.
Right in the middle!
Updating Uncertainty¶
After combining prediction and measurement, the uncertainty also updates!
Formula¶
Example¶
Prediction uncertainty: 5 meters K = 0.2
The uncertainty decreased! We're more certain after combining information.
Key Insight¶
Independent, unbiased measurements with correctly specified variances can reduce uncertainty. Correlation and bias require a different model; duplicating one observation does not create new information. See lesson 12.
Even if both sources are uncertain, combining them makes you more certain.
Practice Problems¶
Problem 1: Simple Weighted Average¶
You have three measurements: - A: 50 (weight = 1) - B: 60 (weight = 2) - C: 55 (weight = 3)
Calculate the weighted average.
Problem 2: From Uncertainty to Weights¶
Two measurements: - Measurement 1: 100 ± 5 - Measurement 2: 110 ± 10
a) Calculate the weight for each measurement b) Calculate the weighted average c) Which measurement had more influence? Why?
Problem 3: Kalman Gain¶
Prediction: 75 ± 3 Measurement: 80 ± 6
a) Calculate the Kalman Gain (K) b) Calculate the new estimate c) Calculate the new uncertainty d) Did the uncertainty increase or decrease?
Problem 4: Extreme Cases¶
For each case, predict whether K will be close to 0, 0.5, or 1:
a) Prediction: 50 ± 1, Measurement: 60 ± 20 b) Prediction: 50 ± 20, Measurement: 60 ± 1 c) Prediction: 50 ± 5, Measurement: 60 ± 5
Problem 5: Real World¶
You're tracking a drone: - Your physics model predicts: altitude = 100 feet ± 2 feet - Barometer measures: altitude = 95 feet ± 8 feet
a) Calculate K b) What's your best estimate of the altitude? c) What's the new uncertainty?
Real World Applications¶
GPS + Inertial Navigation¶
Your phone combines: - GPS: Accurate but slow updates (1 Hz) - Accelerometer: Fast but drifts over time (100 Hz)
The Kalman filter uses weighted averages to combine both!
Robot Localization¶
A robot uses: - Wheel encoders: Measure how far wheels turned - Laser scanner: Measures distance to walls
Both have uncertainty. Weighted average gives best position estimate!
Weather Forecasting¶
Meteorologists combine: - Computer models: Predict weather - Actual measurements: From weather stations
Weighted average gives the forecast you see!
Try It Yourself!¶
Experiment 1: Weighted Coin Flip¶
- Flip a coin 10 times, count heads
- Flip a different coin 100 times, count heads
- Which result do you trust more for estimating probability of heads?
- Calculate a weighted average (weight by number of flips)
Experiment 2: Measuring with Different Tools¶
- Measure something with a ruler (5 times)
- Measure the same thing with a tape measure (5 times)
- Calculate average and standard deviation for each
- Calculate weights based on standard deviations
- Calculate weighted average of the two averages
Experiment 3: Prediction vs Measurement¶
- Drop a ball from a height
- Predict how long it will take to hit the ground (use physics: t = √(2h/g))
- Measure the actual time with a stopwatch
- Estimate uncertainty for each
- Calculate weighted average
Key Concepts¶
- Weighted averages give more importance to more reliable information
- Weight = 1 ÷ Variance (lower uncertainty = higher weight)
- Kalman Gain determines how much to trust measurement vs prediction
- Combining information reduces uncertainty
- The Kalman filter is essentially a smart weighted average
What's Next?¶
Now that we understand weighted averages, the next chapter will teach us about prediction - how to estimate where things will be in the future based on where they are now!
Key Vocabulary - Weighted Average: Average where some values count more than others - Weight: How much importance to give a value - Kalman Gain: The weight given to the measurement in Kalman filter - Variance: Uncertainty squared - Update: Combining prediction and measurement to get new estimate