/
alertReward.py
208 lines (164 loc) · 9.83 KB
/
alertReward.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
# Copyright (c) 2019-2023, RTE (https://www.rte-france.com)
# See AUTHORS.txt
# This Source Code Form is subject to the terms of the Mozilla Public License, version 2.0.
# If a copy of the Mozilla Public License, version 2.0 was not distributed with this file,
# you can obtain one at http://mozilla.org/MPL/2.0/.
# SPDX-License-Identifier: MPL-2.0
# This file is part of Grid2Op, Grid2Op a testbed platform to model sequential decision making in power systems.
import numpy as np
from typing import Optional
from grid2op.Reward.baseReward import BaseReward
from grid2op.dtypes import dt_float, dt_bool, dt_int
class AlertReward(BaseReward):
"""
.. note::
DOC IN PROGRESS !
This reward is based on the "alert feature" where the agent is asked to send information about potential line overload issue
on the grid after unpredictable powerline disconnection (attack of the opponent). The alerts are assessed once per attack.
This rewards is computed as followed:
- if an attack occurs and the agent survives `env.parameters.ALERT_TIME_WINDOW` steps then:
- if the agent sent an alert BEFORE the attack, reward returns `reward_min_no_blackout` (-1 by default)
- if the agent did not sent an alert BEFORE the attack, reward returns `reward_max_no_blackout` (1 by default)
- if an attack occurs and the agent "games over" withing `env.parameters.ALERT_TIME_WINDOW` steps then:
- if the agent sent an alert BEFORE the attack, reward returns `reward_max_blackout` (2 by default)
- if the agent did not sent an alert BEFORE the attack, reward returns `reward_min_blackout` (-10 by default)
- whatever the attacks / no attacks / alert / no alert, if the scenario is completed until the end,
then agent receive `reward_end_episode_bonus` (1 by default)
In all other cases, including but not limited to:
- agent games over but there has been no attack within the previous `env.parameters.ALERT_TIME_WINDOW` (12) steps
- there is no attack
The reward outputs 0.
This is then a "delayed reward": you receive the reward (in general) `env.parameters.ALERT_TIME_WINDOW` after
having sent the alert.
This is also a "sparse reward": in the vast majority of cases it's 0. It is only non zero in case of blackout (at
most once per episode) and each time an attack occurs (and in general there is relatively few attacks)
TODO explain a bit more in the "multi lines attacked"
.. seealso:: :ref:`grid2op-alert-module` section of the doc for more information
Examples
---------
You can use this reward in any environment with:
.. code-block:
import grid2op
from grid2op.Reward import AlertReward
# then you create your environment with it:
# at time of writing, the only env supporting it is "l2rpn_idf_2023"
NAME_OF_THE_ENVIRONMENT = "l2rpn_idf_2023"
env = grid2op.make(NAME_OF_THE_ENVIRONMENT, reward_class=AlertReward)
# and do a step with a "do nothing" action
obs = env.reset()
obs, reward, done, info = env.step(env.action_space())
# the reward is computed with the AlertReward class
"""
def __init__(self,
logger=None,
reward_min_no_blackout=-1.0,
reward_min_blackout=-10.0,
reward_max_no_blackout=1.0,
reward_max_blackout=2.0,
reward_end_episode_bonus=1.0):
BaseReward.__init__(self, logger=logger)
self.reward_min_no_blackout : float = dt_float(reward_min_no_blackout)
self.reward_min_blackout : float = dt_float(reward_min_blackout)
self.reward_max_no_blackout : float = dt_float(reward_max_no_blackout)
self.reward_max_blackout : float = dt_float(reward_max_blackout)
self.reward_end_episode_bonus : float = dt_float(reward_end_episode_bonus)
self.reward_no_game_over : float = dt_float(0.0)
self._reward_range_blackout : float = (self.reward_max_blackout - self.reward_min_blackout)
self.total_time_steps : Optional[int] = dt_int(0)
self.time_window : Optional[int] = None
self._ts_attack : Optional[np.ndarray] = None
self._current_id : int = 0
self._lines_currently_attacked : Optional[np.ndarray] = None
self._alert_launched : Optional[np.ndarray] = None
self._nrows_array : Optional[int] = None
self._i_am_simulate : bool = False
def initialize(self, env: "grid2op.Environment.BaseEnv"):
self.total_time_steps = env.max_episode_duration()
self.time_window = env.parameters.ALERT_TIME_WINDOW
self._nrows_array = self.time_window + 2
self._ts_attack = np.full((self._nrows_array, type(env).dim_alerts), False, dtype=dt_bool)
self._alert_launched = np.full((self._nrows_array, type(env).dim_alerts), False, dtype=dt_bool)
self._current_id = 0
self._lines_currently_attacked = np.full(type(env).dim_alerts, False, dtype=dt_bool)
self._i_am_simulate = self.is_simulated_env(env)
return super().initialize(env)
def reset(self, env):
self._ts_attack[:,:] = False
self._alert_launched[:,:] = False
self._current_id = 0
self._lines_currently_attacked[:] = False
self._i_am_simulate = self.is_simulated_env(env)
return super().reset(env)
def _update_attack(self, env):
if env.infos["opponent_attack_line"] is None:
# no attack at this step
self._lines_currently_attacked[:] = False
self._ts_attack[self._current_id, :] = False
else:
# an attack at this step
lines_attacked = env.infos["opponent_attack_line"][type(env).alertable_line_ids]
# compute the list of lines that are "newly" attacked
new_lines_attacked = lines_attacked & (~self._lines_currently_attacked)
# remember the steps where these lines are attacked
self._ts_attack[self._current_id, new_lines_attacked] = True
# and now update the state of lines under attack
self._lines_currently_attacked[:] = False
self._lines_currently_attacked[lines_attacked] = True
def _update_alert(self, action):
self._alert_launched[self._current_id, :] = 1 * action.raise_alert
def _update_state(self, env, action):
self._current_id += 1
self._current_id %= self._nrows_array
# update attack
self._update_attack(env)
# update alerts
self._update_alert(action)
# update internal state of the environment
# (this is updated in case the reward returns non 0)
env._was_alert_used_after_attack[:] = 0
def _compute_score_attack_blackout(self, env, ts_attack_in_order, indexes_to_look):
# retrieve the lines that have been attacked in the time window
ts_ind, line_ind = np.where(ts_attack_in_order)
line_first_attack, first_ind_line_attacked = np.unique(line_ind, return_index=True)
ts_first_line_attacked = ts_ind[first_ind_line_attacked]
# now retrieve the array starting at the correct place
ts_first_line_attacked_orig = indexes_to_look[ts_first_line_attacked]
# and now look at the previous step if alerts were send
# prev_ts = (ts_first_line_attacked_orig - 1) % self._nrows_array
prev_ts = ts_first_line_attacked_orig
# update the state of the environment
env._was_alert_used_after_attack[line_first_attack] = self._alert_launched[prev_ts, line_first_attack] * 2 - 1
return np.mean(self._alert_launched[prev_ts, line_first_attack]) * self._reward_range_blackout + self.reward_min_blackout
def __call__(self, action, env, has_error, is_done, is_illegal, is_ambiguous):
# retrieve the alert made by the agent
res = 0.
if self._i_am_simulate:
# does not make sense for simulate
return res
if is_done & (not has_error):
# end of episode, no blackout => reward specific for this case
return self.reward_end_episode_bonus
self._update_state(env, action)
if self.is_in_blackout(has_error, is_done):
# I am in blackout, I need to check for attack in the time window
# if there is no attack, I do nothing
indexes_to_look = (np.arange(-self.time_window, 1) + self._current_id) % self._nrows_array # include current step (hence the np.arange(..., **1**))
ts_attack_in_order = self._ts_attack[indexes_to_look, :]
has_attack = (ts_attack_in_order).any()
if has_attack:
# I need to check the alarm for the attacked lines
res = self._compute_score_attack_blackout(env, ts_attack_in_order, indexes_to_look)
else:
# no blackout: i check the first step in the window before me to see if there is an attack,
index_window = (self._current_id - self.time_window) % self._nrows_array
lines_attack = self._ts_attack[index_window, :]
if lines_attack.any():
# prev_ind = (index_window - 1) % self._nrows_array
# I don't need the "-1" because the action is already BEFORE the observation in the reward.
prev_ind = index_window
alert_send = self._alert_launched[prev_ind, lines_attack]
# update the state of the environment
env._was_alert_used_after_attack[lines_attack] = 1 - alert_send * 2
res = (self.reward_min_no_blackout - self.reward_max_no_blackout) * np.mean(alert_send) + self.reward_max_no_blackout
self._ts_attack[index_window, :] = False # attack has been taken into account we "cancel" it
return res