2022-08-20 14:35:29 +00:00
|
|
|
import logging
|
2022-08-28 17:21:57 +00:00
|
|
|
from pathlib import Path
|
2022-08-24 10:54:02 +00:00
|
|
|
from typing import Any, Dict
|
2022-08-20 14:35:29 +00:00
|
|
|
|
|
|
|
import torch as th
|
2022-08-28 17:21:57 +00:00
|
|
|
|
2022-08-20 14:35:29 +00:00
|
|
|
from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
|
2022-08-28 17:21:57 +00:00
|
|
|
from freqtrade.freqai.RL.Base5ActionRLEnv import Actions, Base5ActionRLEnv, Positions
|
2022-08-20 14:35:29 +00:00
|
|
|
from freqtrade.freqai.RL.BaseReinforcementLearningModel import BaseReinforcementLearningModel
|
2022-08-28 17:21:57 +00:00
|
|
|
|
2022-08-20 14:35:29 +00:00
|
|
|
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
|
|
|
|
|
|
class ReinforcementLearner(BaseReinforcementLearningModel):
|
|
|
|
"""
|
2022-11-26 10:51:08 +00:00
|
|
|
Reinforcement Learning Model prediction model.
|
|
|
|
|
|
|
|
Users can inherit from this class to make their own RL model with custom
|
|
|
|
environment/training controls. Define the file as follows:
|
|
|
|
|
|
|
|
```
|
|
|
|
from freqtrade.freqai.prediction_models.ReinforcementLearner import ReinforcementLearner
|
|
|
|
|
|
|
|
class MyCoolRLModel(ReinforcementLearner):
|
|
|
|
```
|
|
|
|
|
|
|
|
Save the file to `user_data/freqaimodels`, then run it with:
|
|
|
|
|
|
|
|
freqtrade trade --freqaimodel MyCoolRLModel --config config.json --strategy SomeCoolStrat
|
|
|
|
|
|
|
|
Here the users can override any of the functions
|
|
|
|
available in the `IFreqaiModel` inheritance tree. Most importantly for RL, this
|
|
|
|
is where the user overrides `MyRLEnv` (see below), to define custom
|
|
|
|
`calculate_reward()` function, or to override any other parts of the environment.
|
|
|
|
|
|
|
|
This class also allows users to override any other part of the IFreqaiModel tree.
|
|
|
|
For example, the user can override `def fit()` or `def train()` or `def predict()`
|
|
|
|
to take fine-tuned control over these processes.
|
|
|
|
|
|
|
|
Another common override may be `def data_cleaning_predict()` where the user can
|
|
|
|
take fine-tuned control over the data handling pipeline.
|
2022-08-20 14:35:29 +00:00
|
|
|
"""
|
|
|
|
|
2022-09-14 22:46:35 +00:00
|
|
|
def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen, **kwargs):
|
2022-09-23 17:17:27 +00:00
|
|
|
"""
|
|
|
|
User customizable fit method
|
2022-11-13 16:43:52 +00:00
|
|
|
:param data_dictionary: dict = common data dictionary containing all train/test
|
2022-09-23 17:17:27 +00:00
|
|
|
features/labels/weights.
|
2022-11-13 16:43:52 +00:00
|
|
|
:param dk: FreqaiDatakitchen = data kitchen for current pair.
|
|
|
|
:return:
|
|
|
|
model Any = trained model to be used for inference in dry/live/backtesting
|
2022-09-23 17:17:27 +00:00
|
|
|
"""
|
2022-08-20 14:35:29 +00:00
|
|
|
train_df = data_dictionary["train_features"]
|
|
|
|
total_timesteps = self.freqai_info["rl_config"]["train_cycles"] * len(train_df)
|
|
|
|
|
|
|
|
policy_kwargs = dict(activation_fn=th.nn.ReLU,
|
2022-10-08 10:10:38 +00:00
|
|
|
net_arch=self.net_arch)
|
2022-08-24 10:54:02 +00:00
|
|
|
|
2022-08-25 09:46:18 +00:00
|
|
|
if dk.pair not in self.dd.model_dictionary or not self.continual_learning:
|
2022-08-24 10:54:02 +00:00
|
|
|
model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs,
|
2022-08-31 14:50:39 +00:00
|
|
|
tensorboard_log=Path(
|
|
|
|
dk.full_path / "tensorboard" / dk.pair.split('/')[0]),
|
2022-08-24 10:54:02 +00:00
|
|
|
**self.freqai_info['model_training_parameters']
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
logger.info('Continual training activated - starting training from previously '
|
|
|
|
'trained agent.')
|
|
|
|
model = self.dd.model_dictionary[dk.pair]
|
|
|
|
model.set_env(self.train_env)
|
2022-08-20 14:35:29 +00:00
|
|
|
|
|
|
|
model.learn(
|
|
|
|
total_timesteps=int(total_timesteps),
|
|
|
|
callback=self.eval_callback
|
|
|
|
)
|
|
|
|
|
|
|
|
if Path(dk.data_path / "best_model.zip").is_file():
|
|
|
|
logger.info('Callback found a best model.')
|
|
|
|
best_model = self.MODELCLASS.load(dk.data_path / "best_model")
|
|
|
|
return best_model
|
|
|
|
|
|
|
|
logger.info('Couldnt find best model, using final model instead.')
|
|
|
|
|
|
|
|
return model
|
|
|
|
|
2022-08-28 17:21:57 +00:00
|
|
|
class MyRLEnv(Base5ActionRLEnv):
|
2022-08-24 16:32:40 +00:00
|
|
|
"""
|
2022-08-25 17:05:51 +00:00
|
|
|
User can override any function in BaseRLEnv and gym.Env. Here the user
|
|
|
|
sets a custom reward based on profit and trade duration.
|
2022-08-24 16:32:40 +00:00
|
|
|
"""
|
2022-08-23 12:58:38 +00:00
|
|
|
|
2022-11-26 11:11:59 +00:00
|
|
|
def calculate_reward(self, action: int) -> float:
|
2022-09-23 17:17:27 +00:00
|
|
|
"""
|
|
|
|
An example reward function. This is the one function that users will likely
|
|
|
|
wish to inject their own creativity into.
|
2022-11-13 16:43:52 +00:00
|
|
|
:param action: int = The action made by the agent for the current candle.
|
|
|
|
:return:
|
2022-09-23 17:17:27 +00:00
|
|
|
float = the reward to give to the agent for current step (used for optimization
|
|
|
|
of weights in NN)
|
|
|
|
"""
|
2022-08-25 17:05:51 +00:00
|
|
|
# first, penalize if the action is not valid
|
|
|
|
if not self._is_valid(action):
|
|
|
|
return -2
|
|
|
|
|
|
|
|
pnl = self.get_unrealized_profit()
|
2022-11-26 11:11:59 +00:00
|
|
|
factor = 100.
|
2022-08-25 17:05:51 +00:00
|
|
|
|
|
|
|
# reward agent for entering trades
|
2022-09-14 22:46:35 +00:00
|
|
|
if (action in (Actions.Long_enter.value, Actions.Short_enter.value)
|
|
|
|
and self._position == Positions.Neutral):
|
2022-08-25 17:05:51 +00:00
|
|
|
return 25
|
|
|
|
# discourage agent from not entering trades
|
|
|
|
if action == Actions.Neutral.value and self._position == Positions.Neutral:
|
|
|
|
return -1
|
|
|
|
|
|
|
|
max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300)
|
2022-11-26 11:11:59 +00:00
|
|
|
trade_duration = self._current_tick - self._last_trade_tick # type: ignore
|
2022-08-25 17:05:51 +00:00
|
|
|
|
|
|
|
if trade_duration <= max_trade_duration:
|
|
|
|
factor *= 1.5
|
|
|
|
elif trade_duration > max_trade_duration:
|
|
|
|
factor *= 0.5
|
|
|
|
|
|
|
|
# discourage sitting in position
|
2022-09-14 22:46:35 +00:00
|
|
|
if (self._position in (Positions.Short, Positions.Long) and
|
|
|
|
action == Actions.Neutral.value):
|
2022-08-25 17:05:51 +00:00
|
|
|
return -1 * trade_duration / max_trade_duration
|
|
|
|
|
|
|
|
# close long
|
|
|
|
if action == Actions.Long_exit.value and self._position == Positions.Long:
|
|
|
|
if pnl > self.profit_aim * self.rr:
|
|
|
|
factor *= self.rl_config['model_reward_parameters'].get('win_reward_factor', 2)
|
2022-10-05 13:58:54 +00:00
|
|
|
return float(pnl * factor)
|
2022-08-25 17:05:51 +00:00
|
|
|
|
|
|
|
# close short
|
|
|
|
if action == Actions.Short_exit.value and self._position == Positions.Short:
|
|
|
|
if pnl > self.profit_aim * self.rr:
|
|
|
|
factor *= self.rl_config['model_reward_parameters'].get('win_reward_factor', 2)
|
2022-10-05 13:58:54 +00:00
|
|
|
return float(pnl * factor)
|
2022-08-25 17:05:51 +00:00
|
|
|
|
|
|
|
return 0.
|