pyliaorachel

Peiyu Liao - Portfolio

A Pen by Rachel on CodePen.

License.

pyliaorachel

# Tested on AWS EC2 Linux

# tree -I '<pattern-to-ignore>' -> display tree
# grep -o '^[^#]*' .gitignore   -> remove comments in .gitignore
# | sed 's/\/$//g'              -> replace trailing '/' after directories
# | tr '\n' '|'                 -> join strings with '|'

tree -I $(grep -o '^[^#]*' .gitignore | sed 's/\/$//g' | tr '\n' '|')

# Tested on MacOSX

pyliaorachel

pip3 uninstall $(basename $(find . -name '*.egg-info') .egg-info)

pyliaorachel

#
# Usage: python3 clean_chinese_corpus.py <corpus-to-clean> > <output>
#


import argparse
import re


# Parse args

pyliaorachel

env = gym.make('CartPole-v0')

# 跑 200 個 episode，每個 episode 都是一次任務嘗試
for i_episode in range(200):
    observation = env.reset() # 讓 environment 重回初始狀態 
    rewards = 0 # 累計各 episode 的 reward 
    for t in range(250): # 設個時限，每個 episode 最多跑 250 個 action
        env.render() # 呈現 environment

        # Key section

pyliaorachel

# 定義 policy
def choose_action(observation):
    pos, v, ang, rot = observation
    return 0 if ang < 0 else 1 # 柱子左傾則小車左移，否則右移 

pyliaorachel

env = gym.make('CartPole-v0')

for i_episode in range(200):
    observation = env.reset()
    rewards = 0
    for t in range(250):
        env.render()

        action = choose_action(observation)
        observation, reward, done, info = env.step(action)

pyliaorachel

def choose_action(state, q_table, action_space, epsilon):
    if np.random.random_sample() < epsilon: # 有 ε 的機率會選擇隨機 action
        return action_space.sample() 
    else: # 其他時間根據現有 policy 選擇 action，也就是在 Q table 裡目前 state 中，選擇擁有最大 Q value 的 action
        return np.argmax(q_table[state]) 

pyliaorachel

def get_state(observation, n_buckets, state_bounds):
    state = [0] * len(observation) 
    for i, s in enumerate(observation): # 每個 feature 有不同的分配
        l, u = state_bounds[i][0], state_bounds[i][1] # 每個 feature 值的範圍上下限
        if s <= l: # 低於下限，分配為 0
            state[i] = 0
        elif s >= u: # 高於上限，分配為最大值
            state[i] = n_buckets[i] - 1
        else: # 範圍內，依比例分配
            state[i] = int(((s - l) / (u - l)) * n_buckets[i])

pyliaorachel

env = gym.make('CartPole-v0')

# 準備 Q table
## Environment 中各個 feature 的 bucket 分配數量
## 1 代表任何值皆表同一 state，也就是這個 feature 其實不重要
n_buckets = (1, 1, 6, 3)

## Action 數量 
n_actions = env.action_space.n