test_flatland_malfunction.py 19 KB
Newer Older
u214892's avatar
u214892 committed
1
import random
2
from typing import Dict, List
u214892's avatar
u214892 committed
3

4
import numpy as np
5
from test_utils import Replay, ReplayConfig, run_replay_config, set_penalties_for_replay
6

7
from flatland.core.env_observation_builder import ObservationBuilder
u214892's avatar
u214892 committed
8
from flatland.core.grid.grid4 import Grid4TransitionsEnum
9
from flatland.core.grid.grid4_utils import get_new_position
u214892's avatar
u214892 committed
10
from flatland.envs.agent_utils import RailAgentStatus
u214892's avatar
u214892 committed
11
12
13
from flatland.envs.rail_env import RailEnv, RailEnvActions
from flatland.envs.rail_generators import complex_rail_generator, sparse_rail_generator
from flatland.envs.schedule_generators import complex_schedule_generator, sparse_schedule_generator
14
15


16
class SingleAgentNavigationObs(ObservationBuilder):
17
    """
18
    We build a representation vector with 3 binary components, indicating which of the 3 available directions
19
20
21
22
23
24
    for each agent (Left, Forward, Right) lead to the shortest path to its target.
    E.g., if taking the Left branch (if available) is the shortest route to the agent's target, the observation vector
    will be [1, 0, 0].
    """

    def __init__(self):
25
        super().__init__()
26
27

    def reset(self):
28
        pass
29

30
    def get(self, handle: int = 0) -> List[int]:
31
32
        agent = self.env.agents[handle]

u214892's avatar
u214892 committed
33
        if agent.status == RailAgentStatus.READY_TO_DEPART:
u214892's avatar
u214892 committed
34
            agent_virtual_position = agent.initial_position
u214892's avatar
u214892 committed
35
        elif agent.status == RailAgentStatus.ACTIVE:
u214892's avatar
u214892 committed
36
            agent_virtual_position = agent.position
u214892's avatar
u214892 committed
37
        elif agent.status == RailAgentStatus.DONE:
u214892's avatar
u214892 committed
38
            agent_virtual_position = agent.target
u214892's avatar
u214892 committed
39
40
41
        else:
            return None

u214892's avatar
u214892 committed
42
        possible_transitions = self.env.rail.get_transitions(*agent_virtual_position, agent.direction)
43
44
45
46
47
48
49
50
51
52
53
        num_transitions = np.count_nonzero(possible_transitions)

        # Start from the current orientation, and see which transitions are available;
        # organize them as [left, forward, right], relative to the current orientation
        # If only one transition is possible, the forward branch is aligned with it.
        if num_transitions == 1:
            observation = [0, 1, 0]
        else:
            min_distances = []
            for direction in [(agent.direction + i) % 4 for i in range(-1, 2)]:
                if possible_transitions[direction]:
u214892's avatar
u214892 committed
54
                    new_position = get_new_position(agent_virtual_position, direction)
u214892's avatar
u214892 committed
55
56
                    min_distances.append(
                        self.env.distance_map.get()[handle, new_position[0], new_position[1], direction])
57
58
59
60
                else:
                    min_distances.append(np.inf)

            observation = [0, 0, 0]
61
            observation[np.argmin(min_distances)] = 1
62
63
64
65
66

        return observation


def test_malfunction_process():
Erik Nygren's avatar
Erik Nygren committed
67
    # Set fixed malfunction duration for this test
68
    stochastic_data = {'prop_malfunction': 1.,
69
                       'malfunction_rate': 1000,
70
                       'min_duration': 3,
Erik Nygren's avatar
Erik Nygren committed
71
                       'max_duration': 3}
72
73
    np.random.seed(5)

74
75
    env = RailEnv(width=20,
                  height=20,
76
77
                  rail_generator=complex_rail_generator(nr_start_goal=10, nr_extra=1, min_dist=5, max_dist=99999,
                                                        seed=0),
78
                  schedule_generator=complex_schedule_generator(),
79
80
81
82
                  number_of_agents=2,
                  obs_builder_object=SingleAgentNavigationObs(),
                  stochastic_data=stochastic_data)

u214892's avatar
u214892 committed
83
    obs = env.reset(False, False, True)
Erik Nygren's avatar
Erik Nygren committed
84
85
86

    # Check that a initial duration for malfunction was assigned
    assert env.agents[0].malfunction_data['next_malfunction'] > 0
u214892's avatar
u214892 committed
87
88
    for agent in env.agents:
        agent.status = RailAgentStatus.ACTIVE
Erik Nygren's avatar
Erik Nygren committed
89

90
    agent_halts = 0
Erik Nygren's avatar
Erik Nygren committed
91
92
    total_down_time = 0
    agent_old_position = env.agents[0].position
93
94
    for step in range(100):
        actions = {}
u214892's avatar
u214892 committed
95

96
97
98
99
        for i in range(len(obs)):
            actions[i] = np.argmax(obs[i]) + 1

        if step % 5 == 0:
Erik Nygren's avatar
Erik Nygren committed
100
            # Stop the agent and set it to be malfunctioning
101
            env.agents[0].malfunction_data['malfunction'] = -1
Erik Nygren's avatar
Erik Nygren committed
102
            env.agents[0].malfunction_data['next_malfunction'] = 0
103
104
            agent_halts += 1

105
106
        obs, all_rewards, done, _ = env.step(actions)

Erik Nygren's avatar
Erik Nygren committed
107
108
109
110
111
112
        if env.agents[0].malfunction_data['malfunction'] > 0:
            agent_malfunctioning = True
        else:
            agent_malfunctioning = False

        if agent_malfunctioning:
Erik Nygren's avatar
Erik Nygren committed
113
            # Check that agent is not moving while malfunctioning
Erik Nygren's avatar
Erik Nygren committed
114
115
116
117
118
            assert agent_old_position == env.agents[0].position

        agent_old_position = env.agents[0].position
        total_down_time += env.agents[0].malfunction_data['malfunction']

Erik Nygren's avatar
Erik Nygren committed
119
    # Check that the appropriate number of malfunctions is achieved
u214892's avatar
u214892 committed
120
121
    assert env.agents[0].malfunction_data['nr_malfunctions'] == 21, "Actual {}".format(
        env.agents[0].malfunction_data['nr_malfunctions'])
Erik Nygren's avatar
Erik Nygren committed
122

Erik Nygren's avatar
Erik Nygren committed
123
    # Check that 20 stops where performed
Erik Nygren's avatar
Erik Nygren committed
124
    assert agent_halts == 20
125

Erik Nygren's avatar
Erik Nygren committed
126
127
    # Check that malfunctioning data was standing around
    assert total_down_time > 0
u214892's avatar
u214892 committed
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145


def test_malfunction_process_statistically():
    """Tests hat malfunctions are produced by stochastic_data!"""
    # Set fixed malfunction duration for this test
    stochastic_data = {'prop_malfunction': 1.,
                       'malfunction_rate': 2,
                       'min_duration': 3,
                       'max_duration': 3}

    env = RailEnv(width=20,
                  height=20,
                  rail_generator=complex_rail_generator(nr_start_goal=10, nr_extra=1, min_dist=5, max_dist=99999,
                                                        seed=0),
                  schedule_generator=complex_schedule_generator(),
                  number_of_agents=2,
                  obs_builder_object=SingleAgentNavigationObs(),
                  stochastic_data=stochastic_data)
u214892's avatar
u214892 committed
146
147
148
    np.random.seed(5)
    random.seed(0)
    env.reset(False, False, True)
u214892's avatar
u214892 committed
149
150
    nb_malfunction = 0
    for step in range(100):
151
        action_dict: Dict[int, RailEnvActions] = {}
u214892's avatar
u214892 committed
152
153
154
155
        for agent in env.agents:
            if agent.malfunction_data['malfunction'] > 0:
                nb_malfunction += 1
            # We randomly select an action
156
            action_dict[agent.handle] = RailEnvActions(np.random.randint(4))
u214892's avatar
u214892 committed
157
158
159
160

        env.step(action_dict)

    # check that generation of malfunctions works as expected
Erik Nygren's avatar
Erik Nygren committed
161
    assert nb_malfunction == 128, "nb_malfunction={}".format(nb_malfunction)
u214892's avatar
u214892 committed
162
163


164
def test_initial_malfunction():
u214892's avatar
u214892 committed
165
166
167
168
169
170
171
172
173
174
    stochastic_data = {'prop_malfunction': 1.,  # Percentage of defective agents
                       'malfunction_rate': 70,  # Rate of malfunction occurence
                       'min_duration': 2,  # Minimal duration of malfunction
                       'max_duration': 5  # Max duration of malfunction
                       }

    speed_ration_map = {1.: 1.,  # Fast passenger train
                        1. / 2.: 0.,  # Fast freight train
                        1. / 3.: 0.,  # Slow commuter train
                        1. / 4.: 0.}  # Slow freight train
u214892's avatar
u214892 committed
175
176
    np.random.seed(5)
    random.seed(0)
u214892's avatar
u214892 committed
177
178
    env = RailEnv(width=25,
                  height=30,
179
                  rail_generator=sparse_rail_generator(max_num_cities=5,
180
181
                                                       max_rails_between_cities=3,
                                                       seed=215545,
182
                                                       grid_mode=True
u214892's avatar
u214892 committed
183
184
185
186
187
                                                       ),
                  schedule_generator=sparse_schedule_generator(speed_ration_map),
                  number_of_agents=1,
                  stochastic_data=stochastic_data,  # Malfunction data generator
                  )
188
    set_penalties_for_replay(env)
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
    replay_config = ReplayConfig(
        replay=[
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                set_malfunction=3,
                malfunction=3,
                reward=env.step_penalty  # full step penalty when malfunctioning
            ),
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=2,
                reward=env.step_penalty  # full step penalty when malfunctioning
            ),
            # malfunction stops in the next step and we're still at the beginning of the cell
            # --> if we take action MOVE_FORWARD, agent should restart and move to the next cell
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=1,
                reward=env.start_penalty + env.step_penalty * 1.0
                # malfunctioning ends: starting and running at speed 1.0
            ),
            Replay(
217
218
                position=(28, 6),
                direction=Grid4TransitionsEnum.EAST,
219
220
221
222
223
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
                reward=env.step_penalty * 1.0  # running at speed 1.0
            ),
            Replay(
224
                position=(27, 6),
225
226
227
228
229
230
231
                direction=Grid4TransitionsEnum.NORTH,
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
                reward=env.step_penalty * 1.0  # running at speed 1.0
            )
        ],
        speed=env.agents[0].speed_data['speed'],
u214892's avatar
u214892 committed
232
233
234
        target=env.agents[0].target,
        initial_position=(28, 5),
        initial_direction=Grid4TransitionsEnum.EAST,
235
    )
236
237
238
    sparse_generator_stable = False
    if sparse_generator_stable:
        run_replay_config(env, [replay_config])
239
240
241


def test_initial_malfunction_stop_moving():
242
243
244
245
246
247
248
249
250
251
252
253
254
    stochastic_data = {'prop_malfunction': 1.,  # Percentage of defective agents
                       'malfunction_rate': 70,  # Rate of malfunction occurence
                       'min_duration': 2,  # Minimal duration of malfunction
                       'max_duration': 5  # Max duration of malfunction
                       }

    speed_ration_map = {1.: 1.,  # Fast passenger train
                        1. / 2.: 0.,  # Fast freight train
                        1. / 3.: 0.,  # Slow commuter train
                        1. / 4.: 0.}  # Slow freight train

    env = RailEnv(width=25,
                  height=30,
255
                  rail_generator=sparse_rail_generator(max_num_cities=5,
256
257
258
                                                       max_rails_between_cities=3,
                                                       seed=215545,
                                                       grid_mode=True
259
260
261
262
263
                                                       ),
                  schedule_generator=sparse_schedule_generator(speed_ration_map),
                  number_of_agents=1,
                  stochastic_data=stochastic_data,  # Malfunction data generator
                  )
264
    set_penalties_for_replay(env)
265
266
267
    replay_config = ReplayConfig(
        replay=[
            Replay(
u214892's avatar
u214892 committed
268
                position=None,
269
                direction=Grid4TransitionsEnum.EAST,
u214892's avatar
u214892 committed
270
                action=RailEnvActions.MOVE_FORWARD,
271
272
                set_malfunction=3,
                malfunction=3,
u214892's avatar
u214892 committed
273
274
                reward=env.step_penalty,  # full step penalty when stopped
                status=RailAgentStatus.READY_TO_DEPART
275
276
277
278
279
280
            ),
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.DO_NOTHING,
                malfunction=2,
u214892's avatar
u214892 committed
281
282
                reward=env.step_penalty,  # full step penalty when stopped
                status=RailAgentStatus.ACTIVE
283
284
285
286
287
288
289
290
291
            ),
            # malfunction stops in the next step and we're still at the beginning of the cell
            # --> if we take action STOP_MOVING, agent should restart without moving
            #
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.STOP_MOVING,
                malfunction=1,
u214892's avatar
u214892 committed
292
293
                reward=env.step_penalty,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
294
295
296
297
298
299
300
            ),
            # we have stopped and do nothing --> should stand still
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.DO_NOTHING,
                malfunction=0,
u214892's avatar
u214892 committed
301
302
                reward=env.step_penalty,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
303
304
305
306
307
308
309
            ),
            # we start to move forward --> should go to next cell now
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
u214892's avatar
u214892 committed
310
311
                reward=env.start_penalty + env.step_penalty * 1.0,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
312
313
            ),
            Replay(
314
315
                position=(28, 6),
                direction=Grid4TransitionsEnum.EAST,
316
317
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
u214892's avatar
u214892 committed
318
319
                reward=env.step_penalty * 1.0,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
320
321
322
            )
        ],
        speed=env.agents[0].speed_data['speed'],
u214892's avatar
u214892 committed
323
324
325
        target=env.agents[0].target,
        initial_position=(28, 5),
        initial_direction=Grid4TransitionsEnum.EAST,
326
    )
327
328
329
    sparse_generator_stable = False
    if sparse_generator_stable:
        run_replay_config(env, [replay_config], activate_agents=False)
330
331


332
def test_initial_malfunction_do_nothing():
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
    random.seed(0)
    np.random.seed(0)

    stochastic_data = {'prop_malfunction': 1.,  # Percentage of defective agents
                       'malfunction_rate': 70,  # Rate of malfunction occurence
                       'min_duration': 2,  # Minimal duration of malfunction
                       'max_duration': 5  # Max duration of malfunction
                       }

    speed_ration_map = {1.: 1.,  # Fast passenger train
                        1. / 2.: 0.,  # Fast freight train
                        1. / 3.: 0.,  # Slow commuter train
                        1. / 4.: 0.}  # Slow freight train

    env = RailEnv(width=25,
                  height=30,
349
                  rail_generator=sparse_rail_generator(max_num_cities=5,
350
351
352
                                                       max_rails_between_cities=3,
                                                       seed=215545,
                                                       grid_mode=True
353
354
355
356
357
                                                       ),
                  schedule_generator=sparse_schedule_generator(speed_ration_map),
                  number_of_agents=1,
                  stochastic_data=stochastic_data,  # Malfunction data generator
                  )
358
    set_penalties_for_replay(env)
359
    replay_config = ReplayConfig(
u214892's avatar
u214892 committed
360
361
362
363
364
365
366
367
368
369
        replay=[
            Replay(
                position=None,
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                set_malfunction=3,
                malfunction=3,
                reward=env.step_penalty,  # full step penalty while malfunctioning
                status=RailAgentStatus.READY_TO_DEPART
            ),
370
371
372
373
374
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.DO_NOTHING,
                malfunction=2,
u214892's avatar
u214892 committed
375
376
                reward=env.step_penalty,  # full step penalty while malfunctioning
                status=RailAgentStatus.ACTIVE
377
378
379
380
381
382
383
384
385
            ),
            # malfunction stops in the next step and we're still at the beginning of the cell
            # --> if we take action DO_NOTHING, agent should restart without moving
            #
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.DO_NOTHING,
                malfunction=1,
u214892's avatar
u214892 committed
386
387
                reward=env.step_penalty,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
388
389
390
391
392
393
394
            ),
            # we haven't started moving yet --> stay here
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.DO_NOTHING,
                malfunction=0,
u214892's avatar
u214892 committed
395
396
                reward=env.step_penalty,  # full step penalty while stopped
                status=RailAgentStatus.ACTIVE
397
398
399
400
401
402
403
            ),
            # we start to move forward --> should go to next cell now
            Replay(
                position=(28, 5),
                direction=Grid4TransitionsEnum.EAST,
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
u214892's avatar
u214892 committed
404
405
                reward=env.start_penalty + env.step_penalty * 1.0,  # start penalty + step penalty for speed 1.0
                status=RailAgentStatus.ACTIVE
406
407
            ),
            Replay(
408
409
                position=(28, 6),
                direction=Grid4TransitionsEnum.EAST,
410
411
                action=RailEnvActions.MOVE_FORWARD,
                malfunction=0,
u214892's avatar
u214892 committed
412
413
                reward=env.step_penalty * 1.0,  # step penalty for speed 1.0
                status=RailAgentStatus.ACTIVE
414
415
416
            )
        ],
        speed=env.agents[0].speed_data['speed'],
u214892's avatar
u214892 committed
417
418
419
        target=env.agents[0].target,
        initial_position=(28, 5),
        initial_direction=Grid4TransitionsEnum.EAST,
420
    )
421
422
423
    sparse_generator_stable = False
    if sparse_generator_stable:
        run_replay_config(env, [replay_config], activate_agents=False)
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442


def test_initial_nextmalfunction_not_below_zero():
    random.seed(0)
    np.random.seed(0)

    stochastic_data = {'prop_malfunction': 1.,  # Percentage of defective agents
                       'malfunction_rate': 0.5,  # Rate of malfunction occurence
                       'min_duration': 5,  # Minimal duration of malfunction
                       'max_duration': 5  # Max duration of malfunction
                       }

    speed_ration_map = {1.: 1.,  # Fast passenger train
                        1. / 2.: 0.,  # Fast freight train
                        1. / 3.: 0.,  # Slow commuter train
                        1. / 4.: 0.}  # Slow freight train

    env = RailEnv(width=25,
                  height=30,
443
                  rail_generator=sparse_rail_generator(max_num_cities=5,
444
445
446
                                                       max_rails_between_cities=3,
                                                       seed=215545,
                                                       grid_mode=True
447
448
449
450
451
452
453
454
455
456
                                                       ),
                  schedule_generator=sparse_schedule_generator(speed_ration_map),
                  number_of_agents=1,
                  stochastic_data=stochastic_data,  # Malfunction data generator
                  )
    agent = env.agents[0]
    env.step({})
    # was next_malfunction was -1 befor the bugfix https://gitlab.aicrowd.com/flatland/flatland/issues/186
    assert agent.malfunction_data['next_malfunction'] >= 0, \
        "next_malfunction should be >=0, found {}".format(agent.malfunction_data['next_malfunction'])