Improving previous experiments

This week, the previous experiment results are improved using new approaches.

Experiments:

• Results for non-LSTM networks with extreme data
• Results for LSTM networks with patched dataset
• Results for best networks in new Montmelo circuit

1. Results for non-LSTM networks with extreme data

In the previous blog, we saw that PilotNet and TinyPilotNet performance was ok and we improved their results with some small tricks. This week, we have tried to get rid of those tricks and instead go a step backward and focus on the dataset. We have decided to train the non-LSTM networks with the full dataset instead of separating it as sequences, shuffling the images on the preprocess part and adding twice to the dataset some extreme cases. These extreme cases are examples where the W is high, looking for the network to focus more on curves.

Below are the results for these experiments, comparing them with last’s weeks results for PilotNet and TinyPilotNet. On the right side, the new experiments are located. In the middle, last’s weeks experiments and on the left hand side the explicit brain results.

As we can see, the improvements is clearly visible both for PilotNet and TinyPilotnet. They can now manage to complete the lap several times.

Click on the images to expand them.

Completed distance

Completed percentage

Circuit diameter

Average speed

Lap seconds

2. Results for LSTM networks with patched dataset

In new experiment the LSTM networks are trained with images without the red line, changing the color of those pixels to match the rest of the floor. Below we can check some examples.

If we look at the result of this experiment, we can say that the deepest still manages to learn how to drive but the Tinypilonet does not perform well enough. The right part of the results is again the new experiment, the middle the previous one and the left is the explicit brain results.

Completed distance

Completed percentage

Circuit diameter

Average speed

Lap seconds

Results for best networks in new Montmeló circuit

A new circuit is added for the experiments, a simulation of the real F1 circuit Montmeló.

Circuit

This is a circuit that is not present in the dataset so it can be used for real test. The many_curves circuit is part of the dataset and the simple circuit is not directly part but a more complex design based on it is, so they both can be biased.

Montmeló circuit has some particularities, having some very difficult turns. Below are the results for all the best networks in this new circuit. None of the trained networks completes the circuit. The deepest LSTM is the one performing the best, completing some part of the circuit.

Completed distance

Completed percentage

Circuit diameter

Average speed

Lap seconds