def make_env(scenario_name, benchmark=False):
'''
Creates a PersonalAgentEnv object as env. This can be used similar to a gym
environment by calling env.reset() and env.step().
Use env.render() to view the environment on the screen.
Input:
scenario_name : name of the scenario from ./scenarios/ to be Returns
(without the .py extension)
benchmark : whether you want to produce benchmarking data
(usually only done during evaluation)
Some useful env properties (see environment.py):
.observation_space : Returns the observation space for each agent
.action_space : Returns the action space for each agent
.n : Returns the number of Agents
'''
from particles.environment import PersonalAgentEnv
import particles.scenarios as scenarios
# load scenario from script
scenario = scenarios.load(scenario_name + ".py").Scenario()
# create world
world = scenario.make_world()
# create multiagent environment
if benchmark:
env = PersonalAgentEnv(world, scenario.reset_world, scenario.reward,
scenario.observation, scenario.benchmark_data)
else:
env = PersonalAgentEnv(world, scenario.reset_world, scenario.reward,
scenario.observation)
return env
if args.specific_agents != '':
specific_agents = args.specific_agents.split(' ')
else:
specific_agents = None
scenario = scenarios.load(args.scenario).Scenario(
kind=args.personalization, num_agents=args.num_agents, seed=args.seed,
load_agents=load_agents, save_agents=None,
specific_agents=specific_agents)
# create world
world = scenario.make_world()
world.episode_len = args.episode_len
env = PersonalAgentEnv(world, scenario.reset_world, scenario.reward,
scenario.observation, info_callback=None,
done_callback=scenario.done, shared_viewer=True)
env.discrete_action_input = True
env.render()
policies = [Reinforce(i, env.observation_space[i].shape[0],
env.action_space[0].n) for i in range(env.n)]
policies[0].load_state_dict(torch.load(
'./trained_models/' + args.trained_model))
eps = np.finfo(np.float32).eps.item()
obs_n = env.reset()
running_reward = 10
support_agents = args.specific_agents.split(' ')
# for agent in support_agents:
# scenario = scenarios.load(args.scenario).Scenario(
# kind=args.personalization, num_agents=args.num_agents, seed=args.seed,
# load_agents=load_agents)
scenario = scenarios.load(args.scenario).Scenario(
kind=args.personalization, num_agents=args.num_agents, seed=args.seed,
load_agents=load_agents, specific_agents=support_agents)
world = scenario.make_world()
world.episode_len = args.episode_len
env = PersonalAgentEnv(world, scenario.reset_world, scenario.reward,
scenario.observation, info_callback=None,
done_callback=scenario.done, shared_viewer=True)
env.discrete_action_input = True
env.seed(args.seed)
policies = [model(i, env.observation_space[i].shape[0],
env.action_space[0].n) for i in range(env.n)]
if args.optimizer == 'Adam':
optimizer = optim.Adam(policies[0].parameters(), lr=args.lr)
elif args.optimizer == 'SGD':
optimizer = optim.SGD(policies[0].parameters(), lr=args.lr)
else:
raise NotImplementedError
scenario.sample_task = True # Start off true
# load scenario from script
scenario = scenarios.load(args.scenario).Scenario(
kind=args.personalization,
num_agents=args.num_agents,
seed=args.seed,
load_agents=None,
save_agents=None,
specific_agents=None)
# create world
world = scenario.make_world()
world.episode_len = args.episode_len
# create multiagent environment
env = PersonalAgentEnv(world,
scenario.reset_world,
scenario.reward,
scenario.observation,
info_callback=None,
done_callback=scenario.done,
shared_viewer=True)
# render call to create viewer window (necessary only for interactive policies)
env.render()
# create interactive policies for each agent
policies = [InteractivePolicy(env, i) for i in range(env.n)]
# execution loop
obs_n = env.reset()
for n in range(100):
t = 0
env.reset()
while t < args.episode_len:
ep_reward = 0
# query for action from each agent's policy