-
Notifications
You must be signed in to change notification settings - Fork 212
Bitcoin Price Series
In the Markov setting of reinforcement learning, it is assumed that the current state of the agent is sufficient to determine the next optimal action of the agent.
In this reinforcement learning setup, the state of the agent is partially observable, that is, there is no explicitly defined state of the agent contrary to the example of chess, where the current state of the board is sufficient to determine the next optimal move for a agent.
As a result of which, the state of the agent is to be inferred from what the agent has observed so far or from historical information.
In this regard, the historical prices series of Bitcoin is taken as input to determine the current state of the network and Deep Sense is used to generate the state representation of the time series.
A time horizon is fixed (depends upon the frequency and the number of entries in the time series) which is the number of entries of historical prices taken from the current timestamp to generate the representation of the state. That is, at any timestamp, the number of previous Bitcoin prices taken to infer the state of the agent is fixed and this time series is input to Deep Sense to generate the state representation.
Along with the historical prices, as inspired from this Medium post, common technical indicators like Simple Moving Average (15 to 60 periods), Relative Strength Index and Average True Range are also input to Deep Sense.
TA-Lib is used for generating these technical indicators.