I'm sorry, but from your statements it looks like you're not really a machine learning developper.
A trained model size is irrelevant of the size of the training data, so I'm curious what do you mean by 'file sizes increase day by day.'. If you store history to re-train your model, obviously the history would grow up, but it would just affect the time it would take when you retrain your model, and should not affect current used model.
also, live training of your model is called reinforcement learniing, and I don't think you need to store any data for this kind of model....
Machine learning is a vast topic to discuss in theory. We can keep on discussing the different types of machine learning like supervised learning, unsupervised learning etc and 'reinforcement learning' is just one type of machine learning based on agent reward which is somewhat like a feedback mechanism.
Yes, we are referring here a type of reinforcement learning along with deep neural network and so we can call it as 'Deep reinforcement learning'.
If you have already developed any machine learning algorithms and trained a model with past data of a couple of months or years and continue to retrain it everyday, then you might understand what we are referring to exactly.
So neither the model size and nor the past trained data remains constant after each training. But we try to disregard the old data as the time passes by, but we have to do it carefully as well to avoid curve fitting and that is what we are referring to as file size keep increasing means the size of the models keep increasing when we add more data for training on every new day.
Most importantly after a threshold point of model size, the system simply freezes and is unable to place any trades properly in time due to massive search space and hence, we have to either use GPU or switch to python to use TPU(Tensor processing unit) used by Google. Now, we are just using some API bridges from MT5 and third party libraries to connect to MT4.
By the way, we are not using any kind of learning in this demo account and this is just a simple neural network with inbuilt weights which we keep changing in the code from time to time. We are referring to machine learning when we use this algo in live accounts only.
If you have any ideas as how to tackle such problems in machine learning and use the algos with least computational power, then you are welcome to discuss here. We will be happy to share some of the most powerful ideas used in latest machine learning algos like Alpha zero used by google deep mind and which can be implemented in forex to develop a real AI which can almost never loose in forex. But everything ends up in freezing the mt4 or mt5 platform when we finally implement the algos and hence, without massive computational power it seems difficult to make it work perfectly in forex live trading though it may work sometimes in demo accounts.
Artificial General Intelligence