Kurzzusammenfassung 9.5.25

Here I testes SAV on the inverted pendulum balance task. As ET I compared the reward distribution of a moving window with one that was sampled at the beginning.

• No retrain gave bad performance after the changes

• Not stop train had problems with collapsing results in the stationary areas (Probably to few samples of bad actions)

• Depending of the size of the change the time to detect was shorter. Large change -> faster detection

• Relearning was triggered and resulted in a bad behavior at the beginning of the training. No transfer was done. So this was expected!