-
Notifications
You must be signed in to change notification settings - Fork 0
/
basic_q_agent.py
191 lines (162 loc) · 7.66 KB
/
basic_q_agent.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
"""
File: basic_q_agent.py
Last modified: 3/5/23 by Peyton
Basic Q-learning with a reduced state space.
Author: Michelle Fu, Peyton Lee
"""
import numpy as np
import pandas
from math import inf
from tqdm import tqdm
from matplotlib import pyplot as plt
from classes import Card, Trick, Player, StateRecord
from baseline_agents import BaselineAgent, GreedyBaseline
from play import play
from evaluate import evaluate
NUM_TRAINING_PASSES = 1
NUM_TRAINING_GAMES = 10000
DEBUGGING = True
# Adapted from project 2.
class QLearner():
def __init__(self, state_space, action_space, discount, alpha):
self.state_space = state_space
self.action_space = action_space
self.discount = discount
self.Q = np.zeros((len(state_space), len(action_space)))
self.alpha = alpha
self.trace = np.zeros((len(state_space), len(action_space)))
# Updates Q based on observation
def update(self, s, a, r, sp):
self.trace[s, a] += 1
self.Q[s, a] += self.alpha * (r + self.discount * np.max(self.Q[sp]) - self.Q[s, a])
# Interpolates values of Q function to unvisited state-action pairs based on 1 nearest neighbor.
def interpolate_Q(self):
print("Interpolating Q...")
for s in tqdm(self.state_space):
for a in self.action_space:
if self.trace[s, a] == 0:
seen, queue = set(), []
closest = self.bfs_Q(s, a, seen, queue)
self.Q[s, a] = closest
self.trace[s, a] += 1
# Helper for finding the nearest neighbor in self.Q.
def bfs_Q(self, s, a, seen, queue):
seen.add((s, a))
queue.append((s, a))
while queue:
curr = queue.pop(0)
if self.trace[curr[0], curr[1]] != 0:
return self.Q[curr[0], curr[1]]
neighbors = [(curr[0]-1, curr[1]), (curr[0]+1, curr[1]), (curr[0], curr[1]-1), (curr[0], curr[1]+1)]
for neighbor in neighbors:
try:
if self.Q[neighbor[0], neighbor[1]] is not None and neighbor not in seen:
seen.add(neighbor)
queue.append(neighbor)
except:
pass
class BasicQAgent(Player):
def __init__(self, pos: int):
super().__init__(pos)
state_space = [i for i in range(4 * 14 ** 4)]
action_space = [i for i in range(52)]
self.q_learner = QLearner(state_space, action_space, discount=1, alpha=0.01)
# needed for plotting
self.win_percent_b = []
self.lose_percent_b = []
self.orw_p_b = []
self.orl_p_b = []
self.win_percent_g = []
self.lose_percent_g = []
self.orw_p_g = []
self.orl_p_g = []
# Employs batch Q-learning with actions constrained by rules
def take_turn(self, trick: Trick, tricks: 'list[Trick]', players: 'list[Player]', hearts_broken: bool) -> Card:
legal_moves = self.get_legal_moves(trick, hearts_broken)
state = self.get_state(trick, tricks, legal_moves)
optimal_card, optimal_Q = None, -inf
for card in legal_moves:
if self.q_learner.Q[state, card.id] > optimal_Q:
optimal_card, optimal_Q = card, self.q_learner.Q[state, card.id]
if DEBUGGING:
hand = sorted([card.name for card in self.hand], key=lambda x: x[-1])
choices = sorted([card.name for card in legal_moves], key=lambda x: x[-1])
print("Hand: ", hand)
print("From ", choices, ", picked ", optimal_card.name)
self.hand.remove(optimal_card)
return optimal_card
# Trains model on a batch of simulated games
def batch_q_learn(self, df: pandas.DataFrame):
for i in range(NUM_TRAINING_PASSES):
print(f"Learning epoch {i}")
for idx, data in tqdm(df.iterrows()):
self.q_learner.update(data["s"], data["a"], data["r"], data["sp"])
# if idx % 5200 == 0:
# self.evaluate_performance()
self.q_learner.interpolate_Q()
def plot_stats(self, type):
plt.figure(1)
x_t = np.array([i * 100 for i in range(len(self.win_percent_b))])
if type == 'baseline':
wins_t = np.array(self.win_percent_b)
losses_t = np.array(self.lose_percent_b)
orw_t = np.array(self.orw_p_b)
orl_t = np.array(self.orl_p_b)
elif type == 'greedy':
wins_t = np.array(self.win_percent_g)
losses_t = np.array(self.lose_percent_g)
orw_t = np.array(self.orw_p_g)
orl_t = np.array(self.orl_p_g)
plt.title('Percentage of wins/losses over time')
plt.xlabel('Episode')
plt.ylabel('Percentage')
plt.plot(x_t, wins_t, label = "wins")
plt.plot(x_t, losses_t, label = "losses")
plt.plot(x_t, orw_t, label = "one round wins")
plt.plot(x_t, orl_t, label = "one round losses")
plt.legend()
plt.savefig("stats" + type + ".png")
plt.clf()
def evaluate_performance(self):
eval_players_b = [self, BaselineAgent(1), BaselineAgent(2), BaselineAgent(3)]
print("Evaluating against baseline...")
one_round_wins, one_round_losses = evaluate(eval_players_b, end_threshold=0, num_evals=5000)
full_game_wins, full_game_losses = evaluate(eval_players_b, end_threshold=100, num_evals=500)
self.win_percent_b.append(full_game_wins[0])
self.lose_percent_b.append(full_game_losses[0])
self.orw_p_b.append(one_round_wins[0])
self.orl_p_b.append(one_round_losses[0])
eval_players_g = [self, GreedyBaseline(1), GreedyBaseline(2), GreedyBaseline(3)]
print("Evaluating against greedy...")
one_round_wins, one_round_losses = evaluate(eval_players_g, end_threshold=0, num_evals=5000)
full_game_wins, full_game_losses = evaluate(eval_players_g, end_threshold=100, num_evals=500)
self.win_percent_g.append(full_game_wins[0])
self.lose_percent_g.append(full_game_losses[0])
self.orw_p_g.append(one_round_wins[0])
self.orl_p_g.append(one_round_losses[0])
# Simulates games and trains the BasicQAgent based on the simulated transition/reward data.
# Saves the training data to a csv. Evaluates the BasicQAgent against random baseline agents.
def main():
state_record = StateRecord()
train_players = [BaselineAgent(0), GreedyBaseline(1), BaselineAgent(2), GreedyBaseline(3)]
print("Generating training data...")
for i in tqdm(range(NUM_TRAINING_GAMES)):
play(train_players, 0, False, state_record)
state_record.write_to_csv('data/train_data.csv')
df = pandas.DataFrame(state_record.record[1:], columns=['player', 's', 'a', 'r', 'sp'])
del state_record
q_agent = BasicQAgent(0)
q_agent.batch_q_learn(df)
eval_players = [q_agent, BaselineAgent(1), BaselineAgent(2), BaselineAgent(3)]
# print("Evaluating...")
# one_round_wins, one_round_losses = evaluate(eval_players, end_threshold=0, num_evals=100000)
# print("One round win percentages: ", one_round_wins)
# print("One round loss percentages: ", one_round_losses)
# full_game_wins, full_game_losses = evaluate(eval_players, end_threshold=100, num_evals=10000)
# print("Full game win percentages: ", full_game_wins)
# print("Full game loss percentages: ", full_game_losses)
evaluate(eval_players, end_threshold=0, num_evals=20, console_game=True)
# q_agent.plot_stats('baseline')
# q_agent.plot_stats('greedy')
if __name__ == '__main__':
main()