-
Notifications
You must be signed in to change notification settings - Fork 0
/
plot_gridworld_agent.py
334 lines (284 loc) · 12.6 KB
/
plot_gridworld_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
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
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
# ruff: noqa: F403, F405
from matplotlib import pyplot as plt
import matplotlib
import numpy as np
import seaborn as sns
import os
import pygame
from src.plot_utils import * # noqa: F405
def plot_agent(critic, env, savepath=None):
"""
Plot and save the Q-tables and the greedy policy of an agent trained on a
Gridworld.
"""
if critic._strategy in ["q_joint", "q_sequential"]:
q_env = critic._q_env._table
q_mon = critic._q_mon._table
filepath = None
if savepath is not None:
filepath = os.path.join(savepath, critic._strategy + "_q_mon.png")
plot_q(q_mon, filepath)
filepath = None
if savepath is not None:
filepath = os.path.join(savepath, critic._strategy + "_q_env.png")
plot_q(q_env[:, None, :, None], filepath)
if critic._strategy in ["q_sequential"]:
n_senv, n_smon, n_aenv, n_amon = q_mon.shape
actions_env = [[a[0] for a in np.argwhere(q_env[i] == q_env[i].max())]
for i in range(n_senv)] # fmt: skip
actions = [[[(a_env, a_mon[0])
for a_env in actions_env[i]
for a_mon in np.argwhere(q_mon[i, j, a_env] == q_mon[i, j, a_env].max())]
for i in range(n_senv)]
for j in range(n_smon)] # fmt: skip
else:
q = q_mon + q_env[:, None, :, None]
n_senv, n_smon, n_aenv, n_amon = q.shape
actions = [[[tuple(a) for a in np.argwhere(q[i, j] == q[i, j].max())]
for i in range(n_senv)]
for j in range(n_smon)] # fmt: skip
else:
q = critic._q_joint._table
filepath = None
if savepath is not None:
filepath = os.path.join(savepath, critic._strategy + "_q_joint.png")
plot_q(q, filepath)
n_senv, n_smon, n_aenv, n_amon = q.shape
actions = [[[tuple(a) for a in np.argwhere(q[i, j] == q[i, j].max())]
for i in range(n_senv)]
for j in range(n_smon)] # fmt: skip
if critic._strategy in ["reward_model", "oracle_with_reward_model"]:
r_env = critic._r_env._table[:, None, :, None]
filepath = None
if savepath is not None:
filepath = os.path.join(savepath, critic._strategy + "_r_env.png")
plot_q(r_env, filepath)
filepath = os.path.join(savepath, critic._strategy + "_policy.png")
plot_pi(env, actions, filepath)
def plot_pi(env, actions_list, filepath=None):
"""
Plot a policy returned for Gridworlds.
actions_list must be a list of actions for every state. For example,
>>> actions[0] <--- monitor state index
[[(3, 0)], <--- list of tuples of (env. action, mon. action) for env. state 0
[(2, 1)], <--- list for env. state 1
[(0, 1), (1, 1)], <--- list for env. state 2
...
Means that the policy:
- In mon. state 0 and env. state 0 executes env. action 3 and mon. action 0,
- In mon. state 0 and env. state 1 executes env. action 2 and mon. action 1,
- In mon. state 0 and env. state 2 executes env. action 0 and mon. action 1,
or env. action 1 and mon. action 1, both with equal probability,
- ... and so on, for all states.
"""
fig, axs = make_subplots(1, len(actions_list), width_per_plot=2, height_per_plot=2)
render_mode = env.unwrapped.render_mode
agent_pos = env.unwrapped.agent_pos
last_pos = env.unwrapped.last_pos
env.unwrapped.render_mode = "rgb_array"
env.unwrapped.agent_pos = None
env.unwrapped.last_pos = None
cell_size = env.unwrapped.cell_size
n_rows, n_cols = env.unwrapped.grid.shape
ORANGE = (255, 175, 0)
WHITE = (255, 255, 255)
for i, actions in enumerate(actions_list):
env.render()
surf = env.unwrapped.window_surface
for y in range(n_rows):
for x in range(n_cols):
pos = (x * cell_size[0], y * cell_size[1])
for a_joint in actions[np.ravel_multi_index((y, x), (n_rows, n_cols))]:
pos = (
x * cell_size[0] + cell_size[0] / 2,
y * cell_size[1] + cell_size[1] / 2,
)
if a_joint[0] == 0: # left
end_pos = (pos[0] - cell_size[0] / 4, pos[1])
arrow_points = (
(end_pos[0], end_pos[1] - cell_size[1] / 5),
(end_pos[0], end_pos[1] + cell_size[1] / 5),
(end_pos[0] - cell_size[0] / 5, end_pos[1]),
)
arrow_width = -(-cell_size[1] // 6)
elif a_joint[0] == 1: # down
end_pos = (pos[0], pos[1] + cell_size[1] / 4)
arrow_points = (
(end_pos[0] - cell_size[0] / 5, end_pos[1]),
(end_pos[0] + cell_size[0] / 5, end_pos[1]),
(end_pos[0], end_pos[1] + cell_size[1] / 5),
)
arrow_width = -(-cell_size[0] // 6)
elif a_joint[0] == 2: # right
end_pos = (pos[0] + cell_size[0] / 4, pos[1])
arrow_points = (
(end_pos[0], end_pos[1] - cell_size[1] / 5),
(end_pos[0], end_pos[1] + cell_size[1] / 5),
(end_pos[0] + cell_size[0] / 5, end_pos[1]),
)
arrow_width = -(-cell_size[1] // 6)
elif a_joint[0] == 3: # up
end_pos = (pos[0], pos[1] - cell_size[1] / 4)
arrow_points = (
(end_pos[0] - cell_size[0] / 5, end_pos[1]),
(end_pos[0] + cell_size[0] / 5, end_pos[1]),
(end_pos[0], end_pos[1] - cell_size[1] / 5),
)
arrow_width = -(-cell_size[0] // 6)
else:
raise ValueError("illegal action")
if a_joint[1] == 0:
pygame.draw.polygon(surf, ORANGE, (pos, end_pos), arrow_width)
pygame.draw.polygon(surf, ORANGE, arrow_points, 0)
elif a_joint[1] == 1:
pygame.draw.polygon(surf, WHITE, (pos, end_pos), 5).scale_by(0.5)
pygame.draw.polygon(surf, WHITE, arrow_points, 5).scale_by(0.5)
else:
raise ValueError("illegal action")
rgb_array = np.transpose(np.array(pygame.surfarray.pixels3d(surf)), (1, 0, 2))
axs[0][i].imshow(rgb_array)
axs[0][i].set_title(f"Monitor State {i}")
axs[0][i].axes.xaxis.set_visible(False)
axs[0][i].axes.yaxis.set_visible(False)
env.unwrapped.render_mode = render_mode
env.unwrapped.agent_pos = agent_pos
env.unwrapped.last_pos = last_pos
if filepath is None:
plt.show()
else:
plt.draw()
plt.savefig(filepath, bbox_inches="tight", pad_inches=0, dpi=1500)
def plot_q_alt(q, filepath=None):
"""
Plot a Q-table of shape (n_env_states, n_mon_states, n_env_actions, n_mon_actions).
The image is a grid of heatmaps where
- Outer rows are monitor states,
- Outer columns are environment states,
- Inner rows are monitor actions,
- Inner columns are environment actions.
Therefore, every heatmap has the value of all (env_action, mon_action) pairs.
The pair with the highest value is highlighted in red.
"""
n_senv, n_smon, n_aenv, n_amon = q.shape
original_q = q.copy()
nonz = q != 0
q[nonz] = np.sign(q[nonz]) * np.log(np.abs(q[nonz]) + 1.0) # for better colormaps
fig, axs = make_subplots(n_smon, n_senv, width_per_plot=n_aenv, height_per_plot=n_amon)
for i in range(n_smon):
for j in range(n_senv):
sns.heatmap(
q[j, i, :, :].T,
ax=axs[i][j],
cmap="plasma",
cbar=False,
annot=original_q[j, i, :, :].T,
fmt=".3f",
vmin=q.min(),
vmax=q.max(),
)
for ii in range(n_amon):
for jj in range(n_aenv):
if q[j, i, jj, ii] == q[j, i, :, :].max():
highlight_cell(axs[i][j], jj, ii, color="r", linewidth=3)
if i == 0:
axs[i][j].set_title(f"Env. State {j}", pad=10)
if j == 0:
axs[i][j].set_ylabel("Mon. Action")
else:
for tick in axs[i][j].yaxis.get_major_ticks():
tick.tick1line.set_visible(False)
tick.tick2line.set_visible(False)
tick.label1.set_visible(False)
tick.label2.set_visible(False)
if i != n_smon - 1:
axs[i][j].xaxis.set_ticks([])
else:
axs[i][j].set_xlabel("Env. Action")
if j == n_senv - 1:
axs[i][j].yaxis.set_label_position("right")
# invisible axis for extra label
ax = fig.add_subplot(n_smon, 1, i + 1, frame_on=False)
ax.tick_params(labelcolor="none", bottom=False, left=False)
ax.set_ylabel(
f"Mon. State {i}",
labelpad=10,
fontsize=matplotlib.rcParams["axes.titlesize"],
)
ax.grid(False)
bbox = axs[i][j].get_position()
rect = matplotlib.patches.Rectangle(
(0, bbox.y0), 1, bbox.height, color="0.7", zorder=-1
)
fig.add_artist(rect)
if filepath is None:
plt.show()
else:
plt.draw()
plt.savefig(filepath, bbox_inches="tight", pad_inches=0, dpi=1500)
def plot_q(q, filepath=None):
"""
Plot a Q-table of shape (n_env_states, n_mon_states, n_env_actions, n_mon_actions).
The image is a grid of heatmaps where
- Outer rows are monitor states,
- Outer columns are monitor actions,
- Inner rows are environment states,
- Inner columns are environment actions.
Therefore, every row across every column has the value of all (env_action, mon_action) pairs.
The pair with the highest value is highlighted in red.
"""
n_senv, n_smon, n_aenv, n_amon = q.shape
original_q = q.copy()
nonz = q != 0
q[nonz] = np.sign(q[nonz]) * np.log(np.abs(q[nonz]) + 1.0) # for better colormaps
fig, axs = make_subplots(n_smon, n_amon, width_per_plot=n_aenv, height_per_plot=n_senv)
for i in range(n_smon):
for j in range(n_amon):
sns.heatmap(
q[:, i, :, j],
ax=axs[i][j],
cmap="plasma",
cbar=False,
annot=original_q[:, i, :, j],
fmt=".3f",
vmin=q.min(),
vmax=q.max(),
)
for ii in range(n_senv):
for jj in range(n_aenv):
if q[ii, i, jj, j] == q[ii, i, :, :].max():
highlight_cell(axs[i][j], jj, ii, color="r", linewidth=2)
if i == 0:
axs[i][j].set_title(f"Mon. Action {j}", pad=10)
if j == 0:
axs[i][j].set_ylabel("Env. State")
else:
for tick in axs[i][j].yaxis.get_major_ticks():
tick.tick1line.set_visible(False)
tick.tick2line.set_visible(False)
tick.label1.set_visible(False)
tick.label2.set_visible(False)
if i != n_smon - 1:
axs[i][j].xaxis.set_ticks([])
else:
axs[i][j].set_xlabel("Env. Action")
if j == 0:
# axs[i][j].yaxis.set_label_position("right")
# invisible axis for extra label
ax = fig.add_subplot(n_smon, 1, i + 1, frame_on=False)
ax.tick_params(labelcolor="none", bottom=False, left=False)
ax.set_ylabel(
f"Mon. State {i}",
labelpad=12,
fontsize=matplotlib.rcParams["axes.titlesize"],
)
ax.grid(False)
bbox = axs[i][j].get_position()
rect = matplotlib.patches.Rectangle(
(0, bbox.y0), 1, bbox.height, color="0.7", zorder=-1
)
fig.add_artist(rect)
if filepath is None:
plt.show()
else:
plt.draw()
plt.savefig(filepath, bbox_inches="tight", pad_inches=0, dpi=1500)