Normally, when you develop a neural network, train it for object recognition (on normal objects like bike, car, plane, dog, cloud, etc.), and it turns out to perform very well, you would like to fine-tune it for e.g. recognizing dog breeds, and this is called fine-tuning.

On the other hand, in reinforcement learning, let's consider a game with rewards on checkpoints $1, 2, 3, \dots n$. When you have a bot that plays and learns using some (e.g. value) neural net, it develops some style of solving the problem to reach some checkpoint $k$, and, after that, when it will reach the $k+1$ checkpoint, it probably will have to revalue the whole strategy.

In this situation, will the bot fine-tune itself? Does it makes sense to keep a replay buffer as is and to "reset" the neural net to train it from scratch, or it's better to stay with fine-tune approach?

If possible, topic-related papers would be very welcome!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.