def __init__(self):
super().__init__()
self.params = defineClass.params()
# action
self.action_df = pd.read_csv(self.params.get('path_class_name'))
actions = self.action_df['clsname'].values.tolist()
self.actions = actions
# actionにindex付け
self.actionIndex = {}
for i, val in enumerate(self.actions):
self.actionIndex[i] = val
# 状態は「心象」「直前のシステム対話行為」「対話の位置」の組み合わせ
self.states_sys_da = ['ct','io','re','qs']
self.states_noun_presence = ['Nx','No']
self.states_impression = ['l','n','h']
self.states = list(itertools.product(self.states_sys_da, self.states_noun_presence, self.states_impression))
self.states = ['_'.join(x) for x in self.states]
# stateにindex付け
self.stateIndex = {}
for i, val in enumerate(self.states):
self.stateIndex[val] = i
self.thres_low_UI = 3
self.thres_high_UI = 5
self.persist_UI_exchgs = 3
self.reward_da_df = pd.read_csv(self.params.get('path_reward_da'), index_col=0)
self.sys_utterance_log = []
self.user_utterance_log = []
self.weight_specific_theme = 0.6
def getUtteranceClassTheme(utterance):
params = defineClass.params()
CLSdf = pd.read_csv(params.get('path_utterance_by_class'))
THEMEdf = pd.read_csv(params.get('path_theme_info'))
if '***' in utterance:
return '-', '-'
else:
clsInfo = CLSdf[CLSdf['agent_utterance'] ==
utterance]['cls'].values.astype('str')
clsInfo = '-'.join(clsInfo)
themeInfo = THEMEdf[THEMEdf['agent_utterance'] ==
utterance]['theme'].values[0]
return clsInfo, themeInfo
else:
print('############\n{}\n############'.format(str(options)))
# seed
np.random.seed(options.seed)
######## python q_learning.py -A [ACT] --model [MODEL] ##############
Qtable_name = '{}/{}_Q'.format(options.model, options.model)
Qfreq_name = '{}/{}_Qfreq'.format(options.model, options.model)
hm_name = '{}/{}_hm.png'.format(options.model, options.model)
reward_name = '{}/{}_reward.png'.format(options.model, options.model)
reward_list_name = '{}/{}_reward.npy'.format(options.model, options.model)
log_name = '{}/{}_log.csv'.format(options.model, options.model)
# params
params = defineClass.params()
# Qを学習
if options.action == 'train':
# dir作成
if not os.path.exists(options.model):
os.mkdir(options.model)
else:
print('model "{}" already exists.'.format(options.model))
if_del = input('### overwrite if you push enter. ###')
env = DialogueEnv()
agent = QlearningAgent(epsilon=options.epsilon)
agent.learn(env,
episode_count=options.n_episode,
learning_rate=options.alpha,
def utterance_selection(user_impression,
user_utterance,
history_utterance,
theme,
candidate_action_num=1):
# params init
params = defineClass.params()
ACTdf = pd.read_csv(params.get('path_utterance_by_class'))
THEMEdf = pd.read_csv(params.get('path_theme_info'))
df = pd.merge(ACTdf[['agent_utterance', 'cls']],
THEMEdf[['agent_utterance', 'theme']],
on='agent_utterance')
# 特定話題の選択に重み（weight）をつける
def weightSpecificTheme(df, w=0.6):
themes = df['theme'].values
themes = [1 - w if t == 'default' else w for t in themes]
themes = [x / np.sum(themes) for x in themes]
df = df.reset_index(drop=True)
select_index = np.random.choice(df.index.values, p=themes)
return df.loc[select_index]
# 1交換目のとき
if user_utterance == None:
next_sysutte = ' *** これから{}の話をしましょう***'.format(theme)
next_sysutte_action = 'change_theme'
history_utterance.add_sysutte(next_sysutte, next_sysutte_action)
return next_sysutte, next_sysutte_action
else:
## ユークリッド距離計算
# クラスごとの代表値を読み込み
feadf = pd.read_csv(params.get('path_feature_average_by_class'))
feadf['prevUI3'] = feadf['UI3average'].values - feadf[
'UI3average_diff'].values
fea_ave = feadf[['prevUI3', 'num_noun_before',
'u_utte_len_before']].values
# 現在の特徴量を計算
num_noun_cnt, u_utte_len_cnt = getNounandLen(user_utterance)
fea_cnt = [[user_impression, num_noun_cnt, u_utte_len_cnt]]
ss = preprocessing.StandardScaler()
ss.fit(fea_ave)
ave_ss = ss.transform(fea_ave)
cnt_ss = ss.transform(fea_cnt)
# 距離を計算
action = np.sort(list(set(df['cls'].values)))
distance = []
for i, val in enumerate(action):
dist = np.linalg.norm(cnt_ss - ave_ss[i])
distance.append(dist)
# 一番近いクラスを選択
Dict = dict(zip(action, distance))
candidate_action = []
for k, v in sorted(Dict.items(),
key=lambda x: x[1])[:candidate_action_num]:
candidate_action.append(k)
n_next_action = np.random.choice(candidate_action)
# 発話の選択
CANDIDATEdf = df[(df['cls'] == n_next_action) & (
(df['theme'] == theme) | (df['theme'] == 'default'))]
CANDIDATEdf = CANDIDATEdf.reset_index(drop=True)
CANDIDATEdf = CANDIDATEdf[['agent_utterance', 'theme', 'cls']]
# 使えないものを削除
for i in range(len(CANDIDATEdf)):
if CANDIDATEdf.loc[i, :][
'agent_utterance'] in history_utterance.history_sysutte:
CANDIDATEdf = CANDIDATEdf.drop(index=[i])
# 候補が残っていないなら，action気にせず候補を決定
if len(CANDIDATEdf) == 0:
CANDIDATEdf = df[(df['theme'] == theme) |
(df['theme'] == 'default')]
CANDIDATEdf = CANDIDATEdf.reset_index(drop=True)
CANDIDATEdf = CANDIDATEdf[['agent_utterance', 'theme', 'cls']]
# 使えないものを削除
for i in range(len(CANDIDATEdf)):
if CANDIDATEdf.loc[i, :][
'agent_utterance'] in history_utterance.history_sysutte:
CANDIDATEdf = CANDIDATEdf.drop(index=[i])
# 選択して終了
SELECTdf = weightSpecificTheme(CANDIDATEdf)
next_sysutte, next_theme, next_action = SELECTdf.values
history_utterance.add_sysutte(next_sysutte, next_action)
return next_sysutte, next_action
