闲言碎语不要表，看看代码就知晓。代码架构如下
# for nature DQN class GPModel(tf.keras.Model): def __init__(self)：... def call(self, inputs, training=None, mask=None):... @staticmethod def __build_dnn_a(activation_function=tf.nn.selu, useBatchNormalization=False):... @staticmethod def __build_dnn_b(activation_function=tf.nn.selu, useBatchNormalization=False):... def _del__(self):... # for dueling DQN class GDModel(tf.keras.Model): def __init__(self):... def call(self, inputs, training=None, mask=None):... @staticmethod def __build_v_net(activation_function=tf.nn.selu, useBatchNormalization=False):... @staticmethod def __build_a_net(activation_function=tf.nn.selu, useBatchNormalization=False):... @staticmethod def __build_dnn_a(activation_function=tf.nn.selu, useBatchNormalization=False):... @staticmethod def __build_dnn_b(activation_function=tf.nn.selu, useBatchNormalization=False):... def _del__(self):...
# Agent class DQNModel(object): def __init__(self, useDuelingDQN=True):... def record(self, s, a, s_, r):... def choose_action(self, state, epsilon=0.1):... def train(self, use_double_dqn = False):... def __del__(self):... def save_weights(self, filepath):... def load_weights(self, filepath):... def get_weights(self):... def set_weights(self, weights):
# Enviroment class SiEnviroment(object): def __init__(self, data):... def reset(self):... def move(self, action):... def last(self):... def describe(self):... def __del__(self):...
# Dog class Dog(object): def __init__(self, profile):... def getStatus(self):... def fit(self):... def evalute(self, verbose=0, sample=0):... def predict(self):... def save_model_weights(self, filepath):... def load_model_weights(self, filepath):...
采用 Dueling DQN + Double DQN运算预测效果如下（部分）
| 资产编号 | 预测动作 | 胜率 | 历史数据测试最大收益率 | 历史数据测试平均收益率 | 历史数据测试最小收益率 |
| 6*****.ss | 观望 | 97.3% | 206.71% | 118.19% | 95.73% |
| 6****6.ss | 卖出 | 45.97% | 158.04% | 106.05% | 85.40% |
| 6****9.ss | 持有 | 81.1% | 174.31% | 112.09% | 94.90% |
| 6***3*.ss | 观望 | 54.64% | 164.89% | 109.20% | 90.80% |
| 6****5.ss | 买入 | 78.56% | 141.74% | 114.12% | 94.11% |
话不多说，说给知者。功能模块可以有偿出售，寻求有兴趣的合作人员，电联：xyz2abc@163.com