-
Notifications
You must be signed in to change notification settings - Fork 70
/
gridworld.py
310 lines (259 loc) · 10.7 KB
/
gridworld.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
import os
import random
import tempfile
from enum import Enum, StrEnum
from collections import defaultdict
from dataclasses import dataclass
from typing import Callable, TypeAlias
import numpy as np
from numpy.testing import assert_almost_equal
from matplotlib import animation
import matplotlib.pyplot as plt
from matplotlib import colors
from IPython.display import Image
class Cell(StrEnum):
START = 'S'
TARGET = 'T'
EMPTY = 'E'
WALL = 'W'
BOMB = 'B'
GLORY = 'G'
NUKE = 'N'
class Grid:
def __init__(self, spec: list[str]):
self.height = len(spec)
self.width = len(spec[0])
self.cells = [[Cell(col) for col in [*row]] for row in spec]
def __str__(self) -> str:
return '\n'.join([' '.join([col.value for col in row]) for row in self.cells]).strip()
def __getitem__(self, key) -> Cell:
x, y = key
return self.cells[self.height - y - 1][x]
class Action(Enum):
UP = 1
RIGHT = 2
DOWN = 3
LEFT = 4
@dataclass(frozen=True)
class State:
x: int = 0
y: int = 0
def pos(self) -> tuple[int]:
return (self.x, self.y)
class GridMDP:
def __init__(self, grid: Grid, start = State(), gamma = 1.0):
self.grid = grid
self.start = start
self.gamma = gamma
self.all_actions = list(Action)
# Arguably, we could filter out the non-reachable states here - but we keep things simple.
self.all_states = [State(x, y) for x in range(0, self.grid.width)
for y in range(0, self.grid.height)]
def is_terminal(self, state: State) -> bool:
cell = self.grid[state.pos()]
return cell in [Cell.TARGET, Cell.GLORY, Cell.BOMB, Cell.NUKE]
def is_reachable(self, n: State) -> bool:
return n.x >= 0 and n.x < self.grid.width \
and n.y >= 0 and n.y < self.grid.height \
and self.grid[n.pos()] != Cell.WALL
def reward(self, state: State, action: Action, next_state: State) -> float:
match self.grid[next_state.pos()]:
case Cell.TARGET:
return 1.0 * self.gamma
case Cell.GLORY:
return 10.0 * self.gamma
case Cell.BOMB:
return -1.0 * self.gamma
case Cell.NUKE:
return -10.0 * self.gamma
case _:
return 0.0
def transition(self, state: State, action: Action, noise = 0.0) -> dict[State, float]:
if not self.is_reachable(state) or self.is_terminal(state):
return {}
def landing(candidate: State) -> State:
return candidate if self.is_reachable(candidate) else state
action_landings = {
Action.UP: landing(State(state.x, state.y + 1)),
Action.RIGHT: landing(State(state.x + 1, state.y)),
Action.DOWN: landing(State(state.x, state.y - 1)),
Action.LEFT: landing(State(state.x - 1, state.y)),
}
mistaken_actions = [Action.UP, Action.DOWN] if action in [Action.LEFT, Action.RIGHT] \
else [Action.LEFT, Action.RIGHT]
next_state = action_landings[action]
mistakes = [action_landings[m] for m in mistaken_actions]
probs = defaultdict(lambda: 0.0)
probs[next_state] = 1.0 - noise
for m in mistakes:
probs[m] += noise / len(mistakes)
assert sum(probs.values()) <= 1.0
assert_almost_equal(sum(probs.values()), 1.0)
return probs
class GridEnv:
def __init__(self, mdp: GridMDP):
self.mdp = mdp
self.state = State()
self.terminated = False
def reset(self) -> State:
self.state = State()
self.terminated = False
return self.state
def step(self, action: Action) -> tuple[State, float, bool]:
if self.terminated:
raise Exception('Environment episode completed, please call reset.')
state_probs = [(s, p) for s, p in self.mdp.transition(self.state, action).items()]
probs = [x[1] for x in state_probs]
next_state_idx = np.random.choice(len(probs), p=probs)
next_state = state_probs[next_state_idx][0]
reward = self.mdp.reward(self.state, action, next_state)
done = self.mdp.is_terminal(next_state)
step = (next_state, reward, done)
self.state = next_state
self.terminated = done
return step
class QTable:
def __init__(self, states: list[State], actions: list[Action]):
self.states = states
self.actions = actions
self.nA = len(actions)
# Why a dict of dict vs. a single dict indexed by tuple (State, Action)?
# Because we want to lookup all action values for a state, and that's more convenient.
self.table: dict[State, dict[Action, float]] = \
{ s : { a : 0.0 for a in actions } for s in states }
def __getitem__(self, key: tuple[State, Action]) -> float:
state, action = key
return self.table[state][action]
def __setitem__(self, key: tuple[State, Action], value: float):
state, action = key
self.table[state][action] = value
def value(self, state: State) -> float:
return max(self.table[state].values())
def best_action(self, state: State) -> Action:
best_action = None
best_v = float('-inf')
actions = list(self.table[state].keys())
random.shuffle(actions) # Random shuffle in case actions have same value...
for a in actions:
v = self[state, a]
if v > best_v:
best_action = a
best_v = v
return best_action
Policy: TypeAlias = Callable[[State], Action]
DEFAULT_GRID = Grid([
'EEET',
'EWEB',
'SEEE',
])
GRID_WORLD_MDP = GridMDP(DEFAULT_GRID, gamma=0.9)
RANDOM_POLICY = lambda _: np.random.choice(list(Action))
# TODO: possibly rename this class. Also action == None means it is terminal...
# Unclear if this is actually a good-enough implementation for the sake of the examples.
@dataclass
class Step:
state: State
action: Action
reward: float
# Keep this outside MDP so we can simulate the user MDP class.
def simulate_mdp(mdp, policy, max_iterations=20) -> list[Step]:
steps = []
state = mdp.start
current_iteration = 0
while current_iteration != max_iterations and \
not mdp.is_terminal(state) and \
mdp.is_reachable(state):
current_iteration += 1
action = policy(state)
state_probs = [(s, p) for s, p in mdp.transition(state, action).items()]
probs = [x[1] for x in state_probs]
next_state = np.random.choice(len(probs), p=probs)
next_state = state_probs[next_state][0]
reward = mdp.reward(state, action, next_state)
steps.append(Step(state, action, reward))
state = next_state
steps.append(Step(state, None, 0.0))
return steps
#--------------------------
# Plotting utilities below
#--------------------------
plt.rcParams["animation.html"] = "jshtml"
plt.rcParams['figure.dpi'] = 100
# plt.ion()
CMAP = colors.ListedColormap(['lavender', 'palegreen', 'white', 'gray', 'lightpink', 'limegreen', 'red'])
CMAP_BOUNDS = [0, 1, 2, 3, 4, 5, 6, 7]
NORM = colors.BoundaryNorm(CMAP_BOUNDS, CMAP.N)
CELL_VALUES = { 'S': 0, 'T': 1, 'E': 2, 'W': 3, 'B': 4, 'G': 5, 'N': 6 }
class PlotData:
def __init__(self, grid, ax, agent_marker, state_value_texts):
self.grid = grid
self.ax = ax
self.agent_marker = agent_marker
self.state_value_texts = state_value_texts
def plot_grid(grid, qtable = None, agent_pos: tuple[int, int] = None):
plt.rcParams["figure.figsize"] = [3.5, 2.5]
# plt.rcParams["figure.autolayout"] = True
colors_matrix = np.array([[CELL_VALUES[marker] for marker in row] for row in grid.cells])
fig, ax = plt.subplots()
ax.imshow(colors_matrix, cmap=CMAP, norm=NORM)
state_value_texts = [[None for _ in range(grid.height)] for _ in range(grid.width)]
if qtable is not None:
for x in range(grid.width):
for y in range(grid.height):
# To make sure there is no conflict with the notebook types.
state = next(filter(lambda s: s.x == x and s.y == y, qtable.table.keys()))
txt = ax.text(x, grid.height - y - 1, f'{qtable.value(state):.2f}',
ha='center', va='center', fontsize=12, color='black')
state_value_texts[x][y] = txt
plt.scatter(0, 2, s=1000, c='white', marker='*')
agent_marker = None if agent_pos is None else \
plt.scatter(agent_pos[0], grid.height - agent_pos[1] -1, s=1000,
c='blue', marker='*', animated=True)
ax.axes.get_xaxis().set_ticks(np.arange(grid.width) + 0.5)
ax.axes.get_yaxis().set_ticks(np.arange(grid.height) + 0.5)
ax.axes.get_xaxis().set_ticklabels([])
ax.axes.get_yaxis().set_ticklabels([])
ax.grid()
# plt.show()
return fig, PlotData(grid, ax, agent_marker, state_value_texts)
def run_simulation(mdp, policy, max_iterations=20, frames_per_state=10):
steps = simulate_mdp(mdp, policy, max_iterations)
# Compute cumulative returns for convenience
returns = [0.0]
for i, s in enumerate(steps):
returns.append(returns[i] + s.reward)
fig, data = plot_grid(mdp.grid, agent_pos=mdp.start.pos())
plt.close(fig)
def animate(frame):
state_frame = frame // frames_per_state
state = steps[state_frame].state
previous_state = None if state_frame == 0 else steps[state_frame - 1].state
stayed_in_place = (state == previous_state)
new_x, new_y = (state.x, mdp.grid.height - state.y - 1)
if frame < frames_per_state:
data.agent_marker.set_offsets((new_x, new_y))
return
old_x, old_y = data.agent_marker.get_offsets()[0]
delta_x = float(new_x - old_x) / 4.
delta_y = float(new_y - old_y) / 4.
anim_frame = frame % frames_per_state
if anim_frame < 4:
if stayed_in_place:
cur_x = new_x + 0.1 * ((-1) ** (anim_frame % 2))
cur_y = new_y
else:
cur_x = old_x + delta_x * anim_frame
cur_y = old_y + delta_y * anim_frame
else:
cur_x = new_x
cur_y = new_y
data.ax.set_title(f'Return: {returns[state_frame]}')
data.agent_marker.set_offsets((cur_x, cur_y))
anim = animation.FuncAnimation(fig, animate, frames=len(steps) * frames_per_state, interval=50)
# Temporary workaround to avoid:
# UserWarning: Animation was deleted without rendering anything.
f = os.path.join(tempfile.tempdir, 'rl_animation.gif')
writergif = animation.PillowWriter(fps=20)
anim.save(f, writer=writergif)
plt.close()
return Image(open(f, 'rb').read())