16. More Data Collection

Recovery Laps

If you drive and record normal laps around the track, even if you record a lot of them, it might not be enough to train your model to drive properly.

Here’s the problem: if your training data is all focused on driving down the middle of the road, your model won’t ever learn what to do if it gets off to the side of the road. And probably when you run your model to predict steering measurements, things won’t go perfectly and the car will wander off to the side of the road at some point.

So you need to teach the car what to do when it’s off on the side of the road.

One approach might be to constantly wander off to the side of the road and then steer back to the middle.

A better approach is to only record data when the car is driving from the side of the road back toward the center line.

So as the human driver, you’re still weaving back and forth between the middle of the road and the shoulder, but you need to turn off data recording when you weave out to the side, and turn it back on when you steer back to the middle.

Driving Counter-Clockwise

Track one has a left turn bias. If you only drive around the first track in a clock-wise direction, the data will be biased towards left turns. One way to combat the bias is to turn the car around and record counter-clockwise laps around the track. Driving counter-clockwise is also like giving the model a new track to learn from, so the model will generalize better.

Using Both Tracks

If you end up using data from only track one, the convolutional neural network could essentially memorize the track. Consider using data from both track one and track two to make a more generalized model.

Collecting Enough Data

How do you know when you have collected enough data? Machine learning involves trying out ideas and testing them to see if they work. If the model is over or underfitting, then try to figure out why and adjust accordingly.

Since this model outputs a single continuous numeric value, one appropriate error metric would be mean squared error. If the mean squared error is high on both a training and validation set, the model is underfitting. If the mean squared error is low on a training set but high on a validation set, the model is overfitting. Collecting more data can help improve a model when the model is overfitting.

What if the model has a low mean squared error on both the training and validation sets, but the car is falling off the track?

Try to figure out the cases where the vehicle is falling off the track. Does it occur only on turns? Then maybe it's important to collect more turning data. The vehicle's driving behavior is only as good as the behavior of the driver who provided the data.

Here are some general guidelines for data collection:

• two or three laps of center lane driving
• one lap of recovery driving from the sides
• one lap focusing on driving smoothly around curves