-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathadr.py
178 lines (155 loc) · 7.51 KB
/
adr.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
import math
from collections import defaultdict
from gym_codecraft.envs.codecraft_vec_env import Rules
class ADR:
def __init__(self,
hstepsize,
stepsize=0.02,
warmup=10,
initial_hardness=0.0,
ruleset: Rules = None,
linear_hardness: bool = False,
max_hardness: float = 200,
hardness_offset: float = 0,
variety: float = 0.7,
step: int = 0,
average_cost_target: float = 0.8):
if ruleset is None:
ruleset = Rules(
cost_modifier_size=4 * [average_cost_target],
cost_modifier_missiles=average_cost_target,
cost_modifier_shields=average_cost_target,
cost_modifier_storage=average_cost_target,
cost_modifier_constructor=average_cost_target,
cost_modifier_engines=average_cost_target,
)
self.ruleset = ruleset
self.variety = variety
self.target_fractions = normalize({b: 1.0 for b in [
'1m', '1s', '1m1p', '2m', '1s1c', '2m1e1p', '3m1p', '2m2p', '2s2c', '2s1c1e', '2s1m1c'
]})
self.target_modifier = average_cost_target
self.stepsize = stepsize
self.warmup = warmup
self.step = step
self.w_ema = 0.5
self.counts = defaultdict(lambda: 0.0)
self.hardness = initial_hardness
self.max_hardness = max_hardness
self.linear_hardness = linear_hardness
self.hardness_offset = hardness_offset
self.stepsize_hardness = hstepsize
self.target_elimination_rate = 0.97
def target_eplenmean(self):
if self.hardness < 25:
return 250 + 6 * self.hardness
elif self.hardness < 50:
return 400 + 4 * (self.hardness - 25)
elif self.hardness < 100:
return 500 + 2 * (self.hardness - 50)
else:
return 600
def adjust(self, counts, elimination_rate, eplenmean, step) -> float:
self.step += 1
stepsize = self.stepsize * min(1.0, self.step / self.warmup)
for build, bfraction in counts.items():
self.counts[build] = (1 - self.w_ema) * bfraction + self.w_ema * self.counts[build]
gradient = defaultdict(lambda: 0.0)
weight = defaultdict(lambda: 0.0)
for build, bfraction in normalize(self.counts).items():
if bfraction == 0:
loss = -100
else:
loss = -self.variety * math.log(self.target_fractions[build] / bfraction)
for module, mfraction in module_norm(build).items():
gradient[module] += mfraction * loss
weight[module] += mfraction
size_key = f'size{size(build)}'
gradient[size_key] += 0.3 * loss
weight[size_key] += 1
for key in gradient.keys():
gradient[key] /= weight[key]
modifier_decay = 1 - self.variety
gradient['m'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_missiles)
gradient['s'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_storage)
gradient['p'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_shields)
gradient['c'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_constructor)
gradient['e'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_engines)
gradient['size1'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_size[0])
gradient['size2'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_size[1])
gradient['size4'] += modifier_decay * math.log(self.target_modifier / self.ruleset.cost_modifier_size[3])
size_weighted_counts = normalize({build: count * size(build) for build, count in self.counts.items()})
average_modifier = 0.0
for build, bfraction in size_weighted_counts.items():
modifier = 0.0
for module, mfraction in module_norm(build).items():
if module == 'm':
modifier += self.ruleset.cost_modifier_missiles * mfraction
if module == 's':
modifier += self.ruleset.cost_modifier_storage * mfraction
if module == 'p':
modifier += self.ruleset.cost_modifier_shields * mfraction
if module == 'c':
modifier += self.ruleset.cost_modifier_constructor * mfraction
if module == 'e':
modifier += self.ruleset.cost_modifier_engines * mfraction
size_modifier = self.ruleset.cost_modifier_size[size(build) - 1]
average_modifier += modifier * size_modifier * bfraction
if average_modifier == 0:
return
average_cost_grad = 10 * math.log(self.target_modifier / average_modifier)
for key, grad in gradient.items():
exponent = stepsize * min(10.0, max(-10.0, grad + average_cost_grad))
multiplier = math.exp(exponent)
if key == 'm':
self.ruleset.cost_modifier_missiles *= multiplier
if key == 's':
self.ruleset.cost_modifier_storage *= multiplier
if key == 'p':
self.ruleset.cost_modifier_shields *= multiplier
if key == 'c':
self.ruleset.cost_modifier_constructor *= multiplier
if key == 'e':
self.ruleset.cost_modifier_engines *= multiplier
if key == 'size1':
self.ruleset.cost_modifier_size[0] *= multiplier
if key == 'size2':
self.ruleset.cost_modifier_size[1] *= multiplier
if key == 'size3':
self.ruleset.cost_modifier_size[2] *= multiplier
if key == 'size4':
self.ruleset.cost_modifier_size[3] *= multiplier
if step > self.hardness_offset:
if self.linear_hardness:
self.hardness = min((step - self.hardness_offset) * self.stepsize_hardness, self.max_hardness)
else:
if eplenmean is not None:
self.hardness += self.stepsize_hardness * (self.target_eplenmean() - eplenmean)
self.hardness = max(0.0, self.hardness)
return average_modifier
def metrics(self):
return {
'adr_missile_cost': self.ruleset.cost_modifier_missiles,
'adr_storage_cost': self.ruleset.cost_modifier_storage,
'adr_constructor_cost': self.ruleset.cost_modifier_constructor,
'adr_engine_cost': self.ruleset.cost_modifier_engines,
'adr_shield_cost': self.ruleset.cost_modifier_shields,
'adr_size1_cost': self.ruleset.cost_modifier_size[0],
'adr_size2_cost': self.ruleset.cost_modifier_size[1],
'adr_size4_cost': self.ruleset.cost_modifier_size[3],
}
def size(build):
modules = 0
for module in [build[i:i+2] for i in range(0, len(build), 2)]:
modules += int(module[:1])
return modules
def module_norm(build):
weights = defaultdict(lambda: 0.0)
for module in [build[i:i+2] for i in range(0, len(build), 2)]:
weights[module[1:]] = float(module[:1])
return normalize(weights)
def normalize(weights):
total = sum(weights.values())
if total == 0:
total = 1
return {key: weight / total for key, weight in weights.items()}