akaniklaus

@laygr Yes although the agent is free to use any type of patterns, in general it acts very similar to mean-reversion based online portfolio selection methods (so it is more...

@yakouyang I would like to ask another question. Do you think it would be meaningful to merge high and low filters to reduce the parameter-size in that respect? Filters are...

Also, do you know anything about hidden sizes which are used in the paper for comparison with the original LSTM? Do you have any idea on how much we should...

Great, thank you!!!

Here is a repository related to that, I am sharing in case it would be useful for you: https://github.com/bkj/pbt

@jiangtaoxie Do you have news about the 1D version yet?

@YZK-yzk Not all coins maybe have the same amount of history available over the exchange (some might not date back to 2015). If you read the paper, those missing rates...

@YZK-yzk I don't remember exactly how it was on the paper but if the decay rate is zero as you have cited, that would mean that they were filled with...

@nimamox Yes, it is possible. From the paper: Note that we use the training objective, as opposed to the validation objective as in Maclaurin et al. (2015), for computing hypergradients....

@nimamox I am trying to have an implementation that check-points the model and optimizer (without hypergradient descent) and then continue training the model with the validation loss and revert back...