stable/freqtrade/freqai/prediction_models/ReinforcementLearner.py

121 lines
4.8 KiB
Python
Raw Normal View History

import logging
from pathlib import Path
from typing import Any, Dict
import numpy as np
import torch as th
from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
from freqtrade.freqai.RL.Base5ActionRLEnv import Actions, Base5ActionRLEnv, Positions
from freqtrade.freqai.RL.BaseReinforcementLearningModel import BaseReinforcementLearningModel
logger = logging.getLogger(__name__)
class ReinforcementLearner(BaseReinforcementLearningModel):
"""
User created Reinforcement Learning Model prediction model.
"""
2022-09-14 22:46:35 +00:00
def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen, **kwargs):
"""
User customizable fit method
:params:
data_dictionary: dict = common data dictionary containing all train/test
features/labels/weights.
dk: FreqaiDatakitchen = data kitchen for current pair.
:returns:
model: Any = trained model to be used for inference in dry/live/backtesting
"""
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,
2022-09-14 22:46:35 +00:00
net_arch=[128, 128])
if dk.pair not in self.dd.model_dictionary or not self.continual_learning:
model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs,
tensorboard_log=Path(
dk.full_path / "tensorboard" / dk.pair.split('/')[0]),
**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)
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
class MyRLEnv(Base5ActionRLEnv):
"""
User can override any function in BaseRLEnv and gym.Env. Here the user
sets a custom reward based on profit and trade duration.
"""
def calculate_reward(self, action):
"""
An example reward function. This is the one function that users will likely
wish to inject their own creativity into.
:params:
action: int = The action made by the agent for the current candle.
:returns:
float = the reward to give to the agent for current step (used for optimization
of weights in NN)
"""
# first, penalize if the action is not valid
if not self._is_valid(action):
return -2
pnl = self.get_unrealized_profit()
rew = np.sign(pnl) * (pnl + 1)
factor = 100
# 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):
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)
trade_duration = self._current_tick - self._last_trade_tick
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):
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)
return float(rew * factor)
# 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)
return float(rew * factor)
return 0.