Как обучить нейросеть торговать на бирже. Обучение с подкреплением

import numpy as np import pandas as pd import random import gym from gym import spaces from tensorflow.keras import Sequential from tensorflow.keras.layers import Dense class CryptoTrader(gym.Env): def __init__(self, data, capital): self.data = data self.capital = capital self.reward_range = (-np.inf, np.inf) self.action_space = spaces.Box(low=-1, high=1, shape=(1,)) self.observation_space = spaces.Box(low=0, high=1, shape=(len(data.columns),)) def reset(self): self.current_step = 0 self.balance = self.capital self.stocks = 0 self.net_worth = self.balance + (self.stocks * self.data.iloc[self.current_step]['Close']) self.profit = 0 return self.data.iloc[self.current_step] def step(self, action): current_price = self.data.iloc[self.current_step]['Close'] done = False if action > 0: # покупка affordable_stocks = int(self.balance / current_price) stocks_bought = int(affordable_stocks * action) cost = stocks_bought * current_price self.balance -= cost self.stocks += stocks_bought elif action < 0: # продажа stocks_sold = int(self.stocks * abs(action)) profit = stocks_sold * current_price self.balance += profit self.stocks -= stocks_sold self.profit += profit self.net_worth = self.balance + (self.stocks * current_price) self.current_step += 1 if self.current_step >= len(self.data.index)-1: done = True obs = self.data.iloc[self.current_step] reward = self.net_worth - self.capital return obs, reward, done, {} def build_model(input_shape): model = Sequential() model.add(Dense(32, activation='relu', input_shape=input_shape)) model.add(Dense(16, activation='relu')) model.add(Dense(1)) model.compile(optimizer='adam', loss='mse') return model def train_model(env, episodes, steps): input_shape = env.observation_space.shape model = build_model(input_shape) for episode in range(episodes): state = env.reset() state = np.reshape(state, [1, input_shape[0]]) for step in range(steps): action = model.predict(state) next_state, reward, done, _ = env.step(action) next_state = np.reshape(next_state, [1, input_shape[0]]) target = reward if not done: target = (reward + 0.95 * np.amax(model.predict(next_state)[0])) target_f = model.predict(state) target_f[0] = target model.fit(state, target_f, epochs=1, verbose=0) state = next_state if done: break print("Episode ", episode + 1, " total net worth: ", env.net_worth) return model if __name__ == '__main__': data = pd.read_csv('data.csv') capital = 1000 env = CryptoTrader(data, capital) model = train_model(env, episodes=100, steps=200)