-
Notifications
You must be signed in to change notification settings - Fork 1
/
rungnnpool.py
180 lines (150 loc) · 5.39 KB
/
rungnnpool.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
from env3rundiv import MultiAgentInvManagementDiv
import gymnasium as gym
from gymnasium.spaces import Dict, Box, Discrete
import numpy as np
from ray.rllib.models import ModelCatalog
import ray
from ray import tune
from ray import air
import os
from ray.rllib.policy.policy import Policy
import time
from ray.rllib.algorithms.ppo import PPOConfig
import json
from ray.rllib.policy.policy import PolicySpec #For policy mapping
from modelpool import GNNActorCriticModelPool
from ray.rllib.algorithms.callbacks import DefaultCallbacks
from ray.rllib.policy.sample_batch import SampleBatch
from ccmodel import FillInActions
ray.shutdown()
def ensure_dir(file_path):
directory = os.path.dirname(file_path)
if not os.path.exists(directory):
os.makedirs(directory)
ModelCatalog.register_custom_model("gnnpool_model", GNNActorCriticModelPool)
#import ray.rllib.algorithms
#from ray.rllib.algorithms.maddpg.maddpg import MADDPGConfig
ray.init()
config = {"connections": {0: [1,2], 1:[3,4], 2:[4, 5], 3:[], 4:[], 5:[]},
#"num_products":2,
"num_nodes": 6}
#num_agents= config["num_nodes"] * config["num_products"]
#num_products = config["num_products"]
num_nodes = config["num_nodes"]
num_agents = num_nodes
def central_critic_observer(agent_obs, **kw):
"""Rewrites the agent obs to include opponent data for training."""
agents = [*agent_obs]
num_agents = len(agents)
obs_space = len(agent_obs[agents[0]])
new_obs = dict()
for agent in agents:
new_obs[agent] = dict()
new_obs[agent]["own_obs"] = agent_obs[agent]
new_obs[agent]["opponent_obs"] = np.zeros((num_agents - 1)*obs_space)
new_obs[agent]["opponent_action"] = np.zeros(2*(num_agents - 1))
i = 0
for other_agent in agents:
if agent != other_agent:
new_obs[agent]["opponent_obs"][i*obs_space:i*obs_space + obs_space] = agent_obs[other_agent]
i += 1
return new_obs
# Test environment
test_env = MultiAgentInvManagementDiv(config)
obs_space = test_env.observation_space
act_space = test_env.action_space
size = obs_space.shape[0]
opponent_obs_space = Box(low=np.tile(obs_space.low, num_agents-1), high=np.tile(obs_space.high, num_agents-1),
dtype=np.float64, shape=(obs_space.shape[0]*(num_agents-1),))
opponent_act_space = Box(low=np.tile(act_space.low, num_agents-1), high=np.tile(act_space.high, num_agents-1),
dtype=np.float64, shape=(act_space.shape[0]*(num_agents-1),))
cc_obs_space = Dict({
"own_obs": obs_space,
"opponent_obs": opponent_obs_space,
"opponent_action": opponent_act_space,
})
print(cc_obs_space)
print("opponent_action", opponent_act_space.shape)
def create_network(connections):
num_nodes = max(connections.keys())
network = np.zeros((num_nodes + 1, num_nodes + 1))
for parent, children in connections.items():
if children:
for child in children:
network[parent][child] = 1
return network
def get_stage(node, network):
reached_root = False
stage = 0
counter = 0
if node == 0:
return 0
while not reached_root:
for i in range(len(network)):
if network[i][node] == 1:
stage += 1
node = i
if node == 0:
return stage
counter += 1
if counter > len(network):
raise Exception("Infinite Loop")
# Agent/Policy ids
agent_ids = []
network = create_network(config["connections"])
echelons = {node: get_stage(node, network) for node in range(len(network))}
agent_ids = []
#agent_ids = [f"{echelons[node]}_{node:02d}_{product}" for node in range(len(network)) for product in range(num_products)]
agent_ids = [f"{echelons[node]}_{node:02d}" for node in range(len(network))]
print(agent_ids)
def policy_dict():
return {f"{agent_id}": PolicySpec() for agent_id in agent_ids}
policy_graphs = {}
for i in range(num_agents):
policy_graphs[agent_ids[i]] = None, cc_obs_space, act_space, {}
# Register environment
def env_creator(config):
return MultiAgentInvManagementDiv(config = config)
tune.register_env("MultiAgentInvManagementDiv", env_creator) # noqa: E501
algo_w_5_policies = (
PPOConfig()
.environment(
env= "MultiAgentInvManagementDiv",
env_config={
"num_agents": num_agents,
},
)
.rollouts(
batch_mode="complete_episodes",
num_rollout_workers=0,
# TODO(avnishn) make a new example compatible w connectors.
enable_connectors=False,)
.callbacks(FillInActions)
.training(
model = {"custom_model": "gnnpool_model",
}
)
.multi_agent(
policies= policy_graphs,
# Map "agent0" -> "pol0", etc...
policy_mapping_fn=(
lambda agent_id, episode, worker, **kwargs: (
print(f"Agent ID: {agent_id}"),
str(agent_id)
)[1]
),
observation_fn = central_critic_observer,
)
.build()
)
iterations = 60
for i in range(iterations):
algo_w_5_policies.train()
path_to_checkpoint = algo_w_5_policies.save()
print(
"An Algorithm checkpoint has been created inside directory: "
f"'{path_to_checkpoint}'. It should contain 5 policies in the 'policies/' sub dir."
)
# Let's terminate the algo for demonstration purposes.
algo_w_5_policies.stop()
print("donee")