CIFAR-10

CIFAR-10 数据集是机器学习领域中一个常用的数据集，主要用于图像分类任务。它包含 60000 张 32x32 彩色图片，分为 10 个类别，每个类别有 6000 张图片。其中，50000 张图片用于训练，10000 张图片用于测试。CIFAR-10 数据集的图像由人工合成，具有固定的分辨率和颜色数量，这使得它成为一个非常适合初学者入门的数据集。
使用 CIFAR-10 数据集进行机器学习的过程如下：

1. 数据预处理：首先，需要对数据进行预处理，例如缩放、数据增强等，以提高模型的泛化能力。
2. 构建模型：选择合适的神经网络模型，例如卷积神经网络（CNN）或循环神经网络（RNN），用于处理图像数据。
3. 训练模型：将预处理后的数据输入到模型中进行训练。在训练过程中，需要不断调整模型的参数，以最小化损失函数。
4. 评估模型：使用测试集评估模型的性能，例如准确率、召回率等指标。
5. 优化模型：根据评估结果，对模型进行优化，例如调整网络结构、学习率等，以提高模型的性能。
6. 应用模型：训练好的模型可以用于实际问题，例如图像分类等。
   总之，CIFAR-10 数据集是一个适合初学者入门的图像数据集，可以用于训练和评估机器学习模型。通过对数据进行预处理，选择合适的模型，并进行充分的训练和优化，可以获得较好的分类效果。

Reinforcement learning
The states are previous history of stock prices, current budget, and current number of shares of a stock.

The actions are buy, sell, or hold (i.e. do nothing).

The stock market data comes from the Yahoo Finance library, pip install yahoo-finance.

%matplotlib inline
from yahoo_finance import Share
from matplotlib import pyplot as plt
import numpy as np
import random
import tensorflow as tf
import random
Define an abstract class called DecisionPolicy:

class DecisionPolicy:
    def select_action(self, current_state, step):
        pass

    def update_q(self, state, action, reward, next_state):
        pass
Here's one way we could implement the decision policy, called a random decision policy:

class RandomDecisionPolicy(DecisionPolicy):
    def __init__(self, actions):
        self.actions = actions

    def select_action(self, current_state, step):
        action = random.choice(self.actions)
        return action
That's a good baseline. Now let's use a smarter approach using a neural network:

class QLearningDecisionPolicy(DecisionPolicy):
    def __init__(self, actions, input_dim):
        self.epsilon = 0.95
        self.gamma = 0.3
        self.actions = actions
        output_dim = len(actions)
        h1_dim = 20

        self.x = tf.placeholder(tf.float32, [None, input_dim])
        self.y = tf.placeholder(tf.float32, [output_dim])
        W1 = tf.Variable(tf.random_normal([input_dim, h1_dim]))
        b1 = tf.Variable(tf.constant(0.1, shape=[h1_dim]))
        h1 = tf.nn.relu(tf.matmul(self.x, W1) + b1)
        W2 = tf.Variable(tf.random_normal([h1_dim, output_dim]))
        b2 = tf.Variable(tf.constant(0.1, shape=[output_dim]))
        self.q = tf.nn.relu(tf.matmul(h1, W2) + b2)

        loss = tf.square(self.y - self.q)
        self.train_op = tf.train.AdamOptimizer(0.001).minimize(loss)
        self.sess = tf.Session()
        self.sess.run(tf.global_variables_initializer())

    def select_action(self, current_state, step):
        threshold = min(self.epsilon, step / 1000.)
        if random.random() < threshold:
            # Exploit best option with probability epsilon
            action_q_vals = self.sess.run(self.q, feed_dict={self.x: current_state})
            action_idx = np.argmax(action_q_vals)  # TODO: replace w/ tensorflow's argmax
            action = self.actions[action_idx]
        else:
            # Explore random option with probability 1 - epsilon
            action = self.actions[random.randint(0, len(self.actions) - 1)]
        return action

    def update_q(self, state, action, reward, next_state):
        action_q_vals = self.sess.run(self.q, feed_dict={self.x: state})
        next_action_q_vals = self.sess.run(self.q, feed_dict={self.x: next_state})
        next_action_idx = np.argmax(next_action_q_vals)
        current_action_idx = self.actions.index(action)
        action_q_vals[0, current_action_idx] = reward + self.gamma * next_action_q_vals[0, next_action_idx]
        action_q_vals = np.squeeze(np.asarray(action_q_vals))
        self.sess.run(self.train_op, feed_dict={self.x: state, self.y: action_q_vals})
Define a function to run a simulation of buying and selling stocks from a market:

def run_simulation(policy, initial_budget, initial_num_stocks, prices, hist, debug=False):
    budget = initial_budget
    num_stocks = initial_num_stocks
    share_value = 0
    transitions = list()
    for i in range(len(prices) - hist - 1):
        if i % 1000 == 0:
            print('progress {:.2f}%'.format(float(100*i) / (len(prices) - hist - 1)))
        current_state = np.asmatrix(np.hstack((prices[i:i+hist], budget, num_stocks)))
        current_portfolio = budget + num_stocks * share_value
        action = policy.select_action(current_state, i)
        share_value = float(prices[i + hist])
        if action == 'Buy' and budget >= share_value:
            budget -= share_value
            num_stocks += 1
        elif action == 'Sell' and num_stocks > 0:
            budget += share_value
            num_stocks -= 1
        else:
            action = 'Hold'
        new_portfolio = budget + num_stocks * share_value
        reward = new_portfolio - current_portfolio
        next_state = np.asmatrix(np.hstack((prices[i+1:i+hist+1], budget, num_stocks)))
        transitions.append((current_state, action, reward, next_state))
        policy.update_q(current_state, action, reward, next_state)

    portfolio = budget + num_stocks * share_value
    if debug:
        print('${}\t{} shares'.format(budget, num_stocks))
    return portfolio
We want to run simulations multiple times and average out the performances:

def run_simulations(policy, budget, num_stocks, prices, hist):
    num_tries = 5
    final_portfolios = list()
    for i in range(num_tries):
        print('Running simulation {}...'.format(i + 1))
        final_portfolio = run_simulation(policy, budget, num_stocks, prices, hist)
        final_portfolios.append(final_portfolio)
        print('Final portfolio: ${}'.format(final_portfolio))
    plt.title('Final Portfolio Value')
    plt.xlabel('Simulation #')
    plt.ylabel('Net worth')
    plt.plot(final_portfolios)
    plt.show()
Call the following function to use the Yahoo Finance library and obtain useful stockmarket data.

def get_prices(share_symbol, start_date, end_date, cache_filename='stock_prices.npy'):
    try:
        stock_prices = np.load(cache_filename)
    except IOError:
        share = Share(share_symbol)
        stock_hist = share.get_historical(start_date, end_date)
        stock_prices = [stock_price['Open'] for stock_price in stock_hist]
        np.save(cache_filename, stock_prices)

    return stock_prices.astype(float)
Who wants to deal with stock market data without looking a pretty plots? No one. So we need this out of law:

def plot_prices(prices):
    plt.title('Opening stock prices')
    plt.xlabel('day')
    plt.ylabel('price ($)')
    plt.plot(prices)
    plt.savefig('prices.png')
    plt.show()
Train a reinforcement learning policy:

if __name__ == '__main__':
    prices = get_prices('MSFT', '1992-07-22', '2016-07-22')
    plot_prices(prices)
    actions = ['Buy', 'Sell', 'Hold']
    hist = 3
    # policy = RandomDecisionPolicy(actions)
    policy = QLearningDecisionPolicy(actions, hist + 2)
    budget = 100000.0
    num_stocks = 0
    run_simulations(policy, budget, num_stocks, prices, hist)

Running simulation 1...
progress 0.00%
progress 16.55%
progress 33.10%
progress 49.64%
progress 66.19%
progress 82.74%
progress 99.29%
Final portfolio: $211337.9906860001
Running simulation 2...
progress 0.00%
progress 16.55%
progress 33.10%
progress 49.64%
progress 66.19%
progress 82.74%
progress 99.29%
Final portfolio: $217982.47320000027
Running simulation 3...
progress 0.00%
progress 16.55%
progress 33.10%
progress 49.64%
progress 66.19%
progress 82.74%
progress 99.29%
Final portfolio: $218393.6024270001
Running simulation 4...
progress 0.00%
progress 16.55%
progress 33.10%
progress 49.64%
progress 66.19%
progress 82.74%
progress 99.29%
Final portfolio: $217917.51057200006
Running simulation 5...
progress 0.00%
progress 16.55%
progress 33.10%
progress 49.64%
progress 66.19%
progress 82.74%
progress 99.29%
Final portfolio: $219461.85500000033