FinRL StockTradingEnv状态表示问题：新增特征后维度不匹配

我正在FinRL元环境中调整集成代理的自定义交易策略，添加了神经网络生成的新特征（MYNN）来捕捉数据集的时间动态。处理后数据集的列信息及状态空间计算如下：

INDICATORS = ['macd','rsi_30','cci_30','dx_30','wr_30','atr_30','chop_30','mfi_30','boll_ub','boll_lb','close_30_sma','close_60_sma']

MYNN = ['NN0','NN1','NN2','NN3','NN4']

df_columns = ['date','open','high','low','close','volume','tic','day','macd','rsi_30','cci_30','dx_30','wr_30','atr_30','chop_30','mfi_30','boll_ub','boll_lb','close_30_sma','close_60_sma','vix','turbulence','NN0','NN1','NN2','NN3','NN4']

stock_dimension = len(df_final.tic.unique())
state_space = 1 + 2*stock_dimension + (len(INDICATORS)+len(MYNN ))*stock_dimension
print(f"Stock Dimension: {stock_dimension}, State Space: {state_space}")
#Stock Dimension: 27, State Space: 514

按公式计算状态空间应为514，但环境的_initiate_state函数生成的状态维度仅为442，导致报错：无法将形状(442,)的输入数组广播到形状(514,)。我已经调整了该函数，但仍无法定位问题，目标是让代理能获取新增特征，确保这些特征参与状态初始化。调整后的函数代码如下：

#new_features_list -----> myNN features
def _initiate_state(self):
    if self.initial:
        # For Initial State
        if len(self.df.tic.unique()) > 1:
            # for multiple stock
            state = (
                [self.initial_amount]
                + self.data.close.values.tolist()
                + self.num_stock_shares
                + sum(
                    (self.data[tech].values.tolist() for tech in self.tech_indicator_list),
                    [],
                )
                + sum((self.data[feature].values.tolist() for feature in self.new_features_list), [])
            )
        else:
            # for single stock
            state = (
                [self.initial_amount]
                + [self.data.close]
                + [0] * self.stock_dim
                + sum(([self.data[tech]] for tech in self.tech_indicator_list), [])
                + sum(([self.data[feature]] for feature in self.new_features_list), [])
            )
    else:
        # Using Previous State
        if len(self.df.tic.unique()) > 1:
            # for multiple stock
            state = (
                [self.previous_state[0]]
                + self.data.close.values.tolist()
                + self.previous_state[(self.stock_dim + 1) : (self.stock_dim * 2 + 1)]
                + sum(
                    (self.data[tech].values.tolist() for tech in self.tech_indicator_list),
                    [],
                )
                + sum((self.data[feature].values.tolist() for feature in self.new_features_list), [])
            )
        else:
            # for single stock
            state = (
                [self.previous_state[0]]
                + [self.data.close]
                + self.previous_state[(self.stock_dim + 1) : (self.stock_dim * 2 + 1)]
                + sum(([self.data[tech]] for tech in self.tech_indicator_list), [])
                + sum(([self.data[feature]] for feature in self.new_features_list), [])
            )
    print(f"Initialized state shape: {len(state)} (Expected: {self.state_space})")
    return state

核心排查方向 #

状态维度差为514-442=72，说明状态构造过程中某部分元素缺失，按以下步骤逐一验证：

1. 确认特征列表的完整性
   在_initiate_state函数开头添加打印，检查实际使用的特征数量是否与定义一致：

   print(f"Tech indicators count: {len(self.tech_indicator_list)}")
   print(f"NN features count: {len(self.new_features_list)}")
   

   确保前者长度为12（对应INDICATORS），后者为5（对应MYNN），如果数量不足，需修正tech_indicator_list和new_features_list的赋值逻辑。

2. 验证状态空间参数已更新
   检查环境初始化代码，确保self.state_space被设置为计算出的514，而非沿用FinRL默认值：

   self.state_space = 1 + 2*stock_dimension + (len(INDICATORS)+len(MYNN ))*stock_dimension
   

3. 检查持股数列表维度
   多股票场景下，self.num_stock_shares应为长度27的列表（初始全为0），如果是单元素列表会直接少26个维度，添加打印验证：

   print(f"Num stock shares length: {len(self.num_stock_shares)}")
   

4. 拆分状态构造过程定位缺失
   将状态拆分为各组成部分并打印长度，精准定位哪一块元素不足：

   # 在多股票初始状态分支添加
   part1 = [self.initial_amount]
   part2 = self.data.close.values.tolist()
   part3 = self.num_stock_shares
   part4 = sum((self.data[tech].values.tolist() for tech in self.tech_indicator_list), [])
   part5 = sum((self.data[feature].values.tolist() for feature in self.new_features_list), [])
   print(f"各部分长度：part1={len(part1)}, part2={len(part2)}, part3={len(part3)}, part4={len(part4)}, part5={len(part5)}")
   state = part1 + part2 + part3 + part4 + part5
   

   正常输出应为：part1=1, part2=27, part3=27, part4=324, part5=135，总和514，哪块长度不符就排查对应数据源或列表逻辑。

内容的提问来源于stack exchange，提问作者alee stvr