Improve the training by randomly setting the starting point

This is an improvement for all kind of RL stuff, the starting point sets in the first third of the data which ensures that the model not only optimizes the policy or Q values to the choosen timerange.

freqtrade/freqai/RL/BaseEnvironment.py

@@ -2,7 +2,7 @@ import logging
 from abc import abstractmethod
 from enum import Enum
 from typing import Optional
-
+import random
 import gym
 import numpy as np
 import pandas as pd
@@ -121,6 +121,9 @@ class BaseEnvironment(gym.Env):
         self._done = False
 
         if self.starting_point is True:
+            length_of_data = int(len(self.prices)/3) # 1/3 looks okay in any case of short or long training timerange 
+            start_tick = random.randint(self.window_size+1, length_of_data)
+            self._start_tick = start_tick
             self._position_history = (self._start_tick * [None]) + [self._position]
         else:
             self._position_history = (self.window_size * [None]) + [self._position]