File size: 9,098 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from collections import namedtuple
import numpy as np

ORIGINAL_AGENT = "me"
OPPONENT_AGENT = "opponent"


class SMACReward:
    info_template = namedtuple('EnvElementInfo', ['shape', 'value', 'to_agent_processor', 'from_agent_processor'])

    def __init__(
        self,
        n_agents,
        n_enemies,
        two_player,
        reward_type,
        max_reward,
        reward_scale=True,
        reduce_agent=True,
        reward_only_positive=True
    ):
        self.reward_only_positive = reward_only_positive
        self.reward_scale = reward_scale
        self.max_reward = max_reward
        self.reward_death_value = 10
        self.reward_win = 200
        self.reward_defeat = 0
        self.reward_negative_scale = 0.5
        self.reward_scale_rate = 20
        self.reduce_agent = reduce_agent
        self.reward_type = reward_type
        assert self.reward_type in ['sparse', 'original', 'new']
        self.n_agents = n_agents
        self.n_enemies = n_enemies

        self.death_tracker_ally = np.zeros(n_agents)
        self.death_tracker_enemy = np.zeros(n_enemies)

        self.two_player = two_player

    def reset(self, max_reward):
        self.max_reward = max_reward
        if self.reward_type == 'original':
            self.info().value['max'] = self.max_reward / self.reward_scale_rate
        self.death_tracker_ally.fill(0)
        self.death_tracker_enemy.fill(0)

    def get_reward(self, engine, action, game_end_code, win_counted, defeat_counted):
        reward = {
            ORIGINAL_AGENT: np.asarray(self.reward_battle_split(engine, action, is_opponent=False)),
            OPPONENT_AGENT: np.asarray(self.reward_battle_split(engine, action, is_opponent=True))
        }
        for k in reward:
            if reward[k].shape == ():
                reward[k] = np.expand_dims(reward[k], 0)

        if game_end_code is not None:
            # Battle is over
            if game_end_code == 1 and not win_counted:
                if self.reward_type != "sparse":
                    reward[ORIGINAL_AGENT] += self.reward_win
                    reward[OPPONENT_AGENT] += self.reward_defeat
                else:
                    reward[ORIGINAL_AGENT] += 1
                    reward[OPPONENT_AGENT] += -1
            elif game_end_code == -1 and not defeat_counted:
                if self.reward_type != "sparse":
                    reward[ORIGINAL_AGENT] += self.reward_defeat
                    reward[OPPONENT_AGENT] += self.reward_win
                else:
                    reward[ORIGINAL_AGENT] += -1
                    reward[OPPONENT_AGENT] += 1
            # Note: if draw happen, the game_end_code may still be None.

        if self.reward_scale:
            # rescale to 0~1
            min_val, max_val = self.info().value['min'], self.info().value['max']
            reward[ORIGINAL_AGENT] = (reward[ORIGINAL_AGENT] - min_val) / (max_val - min_val)
            reward[OPPONENT_AGENT] = (reward[OPPONENT_AGENT] - min_val) / (max_val - min_val)

        return reward

    def reward_battle_split(self, engine, action, is_opponent=False):
        """Reward function when self.reward_type != 'sparse'.
        Returns accumulative hit/shield point damage dealt to the enemy
        + reward_death_value per enemy unit killed, and, in case
        self.reward_only_positive == False, - (damage dealt to ally units
        + reward_death_value per ally unit killed) * self.reward_negative_scale
        """

        num_agents = engine.n_agents if not is_opponent else engine.n_enemies
        num_enmies = engine.n_agents if is_opponent else engine.n_enemies

        if self.reward_type == 'sparse':
            if self.reduce_agent:
                return 0.
            else:
                return np.zeros(num_agents)

        # if self.reward_type != 'original':
        assert self.reward_type == 'original', 'reward_type={} is not supported!'.format(self.reward_type)
        delta_deaths = np.zeros([num_agents])
        reward = np.zeros([num_agents])
        delta_ally = np.zeros([num_agents])
        delta_enemy = np.zeros([num_enmies])
        delta_death_enemy = np.zeros([num_enmies])

        neg_scale = self.reward_negative_scale

        # update deaths
        if is_opponent:
            iterator = engine.enemies.items()
            previous_units = engine.previous_enemy_units
            death_tracker = self.death_tracker_enemy
        else:
            iterator = engine.agents.items()
            previous_units = engine.previous_ally_units
            death_tracker = self.death_tracker_ally

        num_players = 2 if self.two_player else 1
        for al_id, al_unit in iterator:
            if death_tracker[al_id] < num_players:
                # did not die so far
                prev_health = (previous_units[al_id].health + previous_units[al_id].shield)
                if al_unit.health == 0:
                    # just died
                    death_tracker[al_id] += 1
                    delta_deaths[al_id] -= self.reward_death_value * neg_scale
                    delta_ally[al_id] += prev_health * neg_scale
                else:
                    # still alive
                    delta_ally[al_id] += neg_scale * (prev_health - al_unit.health - al_unit.shield)

        # Calculate the damage to opponent.
        if is_opponent:
            iterator = engine.agents.items()
            previous_units = engine.previous_ally_units
            death_tracker = self.death_tracker_ally
        else:
            iterator = engine.enemies.items()
            previous_units = engine.previous_enemy_units
            death_tracker = self.death_tracker_enemy

        for e_id, e_unit in iterator:
            if death_tracker[e_id] < num_players:
                prev_health = (previous_units[e_id].health + previous_units[e_id].shield)
                if e_unit.health == 0:
                    death_tracker[e_id] += 1
                    delta_death_enemy[e_id] += self.reward_death_value
                    delta_enemy[e_id] += prev_health
                else:
                    delta_enemy[e_id] += prev_health - e_unit.health - e_unit.shield
                # if e_unit.health == 0:
                #     death_tracker[e_id] += 1
                #     delta_death_enemy[e_id] += self.reward_death_value
                #     normed_delta_health = prev_health / (e_unit.health_max + e_unit.shield_max)
                #     delta_enemy[e_id] += normed_delta_health * self.reward_death_value
                # else:
                #     normed_delta_health = (prev_health - e_unit.health -
                #                            e_unit.shield) / (e_unit.health_max + e_unit.shield_max)
                #     delta_enemy[e_id] += normed_delta_health * self.reward_death_value

        # if self.reward_type == 'original':
        #     if self.reduce_agent:
        #         total_reward = sum(delta_deaths) + sum(delta_death_enemy) + sum(delta_enemy)
        #         return total_reward
        #     else:
        #         total_reward = sum(delta_deaths) + sum(delta_death_enemy) + sum(delta_enemy) / num_agents
        #         return np.ones(num_agents) * total_reward

        # Attacking reward
        # if isinstance(action, dict):
        #     my_action = action["me"] if not is_opponent else action["opponent"]
        # else:
        #     my_action = action
        # for my_id, my_action in enumerate(my_action):
        #     if my_action > 5:
        #         reward[my_id] += 2

        if self.reward_only_positive:
            # reward = abs((delta_deaths + delta_death_enemy + delta_enemy).sum())
            reward = abs(delta_deaths.sum() + delta_death_enemy.sum() + delta_enemy.sum())
        else:
            reward = delta_deaths.sum() + delta_death_enemy.sum() + delta_enemy.sum() - delta_ally.sum()

        return reward

    def info(self):
        if self.reward_type == 'sparse':
            value = {'min': -1, 'max': 1}
        elif self.reward_type == 'original':
            value = {'min': 0, 'max': self.max_reward / self.reward_scale_rate}
            # value = {'min': 0, 'max': 75.5}
            # value = {'min': 0, 'max': self.max_reward / 75.5}
        #     # TODO(nyz) health + shield range
        #     if self.reduce_agent:
        #         value = {'min': 0, 'max': (self.reward_win + self.reward_death_value * self.n_enemies +1230)/20}
        #     else:
        #         value = {'min': 0, 'max': self.reward_win + self.reward_death_value * self.n_enemies / self.n_agents}
        # elif self.reward_type == 'new':
        #     if self.reduce_agent:
        #         value = {'min': 0, 'max': self.reward_win + 2 + self.reward_death_value * self.n_enemies}
        #     else:
        #         value = {
        #             'min': 0,
        #             'max': self.reward_win + 2 + self.reward_death_value * self.n_enemies / self.n_agents
        #         }
        shape = (1, ) if self.reduce_agent else (self.n_agents, )
        return SMACReward.info_template(shape, value, None, None)