Rationale for training only on 10% of the buffer ? #7
Description
When defining the batch size in training_go.py, you comment : 'To avoid overfitting, we want to make sure the agent only sees ~10% of samples in the replay over one checkpoint.' 'That is, batch_size * ckpt_interval <= replay_capacity * 0.1'. Can you expand on this choice ? Intuitively training on a small sample of the buffer will foster overfitting rather than prevent it doesn't it ? Can you explain more in details this choice please :)