def rom_agent(model_name,
i,
env,
M,
u_range,
base_kwargs,
g=False,
mu=1.5,
game_name='matrix'):
print(model_name)
joint = False
squash = True
opponent_modelling = True
squash_func = tf.tanh
correct_tanh = True
sampling = False
# TODO deal with particle problem.
if 'particle' in game_name:
sampling = True
squash = True
squash_func = tf.nn.softmax
correct_tanh = False
pool = SimpleReplayBuffer(env.env_specs,
max_replay_buffer_size=1e6,
joint=joint,
agent_id=i)
opponent_policy = GaussianPolicy(env.env_specs,
hidden_layer_sizes=(M, M),
squash=True,
joint=False,
agent_id=i,
name='opponent_policy')
conditional_policy = GaussianConditionalPolicy(
env.env_specs,
cond_policy=opponent_policy,
hidden_layer_sizes=(M, M),
name='gaussian_conditional_policy',
opponent_policy=False,
squash=True,
joint=False,
agent_id=i)
with tf.variable_scope('target_levelk_{}'.format(i), reuse=True):
target_opponent_policy = GaussianPolicy(env.env_specs,
hidden_layer_sizes=(M, M),
squash=True,
joint=False,
agent_id=i,
name='target_opponent_policy')
target_conditional_policy = GaussianConditionalPolicy(
env.env_specs,
cond_policy=target_opponent_policy,
hidden_layer_sizes=(M, M),
name='target_gaussian_conditional_policy',
opponent_policy=False,
squash=True,
joint=False,
agent_id=i)
joint_qf = NNJointQFunction(env_spec=env.env_specs,
hidden_layer_sizes=[M, M],
joint=joint,
agent_id=i)
target_joint_qf = NNJointQFunction(env_spec=env.env_specs,
hidden_layer_sizes=[M, M],
name='target_joint_qf',
agent_id=i)
plotter = None
agent = ROMMEO(base_kwargs=base_kwargs,
agent_id=i,
env=env,
pool=pool,
joint_qf=joint_qf,
target_joint_qf=target_joint_qf,
policy=conditional_policy,
opponent_policy=opponent_policy,
target_policy=target_conditional_policy,
plotter=plotter,
policy_lr=1e-2,
qf_lr=1e-2,
joint=True,
value_n_particles=16,
kernel_fn=adaptive_isotropic_gaussian_kernel,
kernel_n_particles=32,
kernel_update_ratio=0.5,
td_target_update_interval=1,
discount=0.95,
reward_scale=1,
tau=0.01,
save_full_state=False,
opponent_modelling=True)
return agent
def pr2ac_agent(model_name,
i,
env,
M,
u_range,
base_kwargs,
k=0,
g=False,
mu=1.5,
game_name='matrix',
aux=True):
joint = False
squash = True
squash_func = tf.tanh
correct_tanh = True
sampling = False
if 'particle' in game_name:
sampling = True
squash = True
squash_func = tf.nn.softmax
correct_tanh = False
pool = SimpleReplayBuffer(env.env_specs,
max_replay_buffer_size=1e6,
joint=joint,
agent_id=i)
opponent_conditional_policy = StochasticNNConditionalPolicy(
env.env_specs,
hidden_layer_sizes=(M, M),
name='opponent_conditional_policy',
squash=squash,
squash_func=squash_func,
sampling=sampling,
u_range=u_range,
joint=joint,
agent_id=i)
if g:
policies = []
target_policies = []
for kk in range(1, k + 1):
policy, target_policy = get_level_k_policy(
env,
kk,
M,
i,
u_range,
opponent_conditional_policy,
game_name=game_name)
policies.append(policy)
target_policies.append(target_policy)
policy = GeneralizedMultiLevelPolicy(env.env_specs,
policies=policies,
agent_id=i,
k=k,
mu=mu)
target_policy = GeneralizedMultiLevelPolicy(env.env_specs,
policies=policies,
agent_id=i,
k=k,
mu=mu,
correct_tanh=correct_tanh)
else:
if k == 0:
policy = DeterministicNNPolicy(env.env_specs,
hidden_layer_sizes=(M, M),
squash=squash,
squash_func=squash_func,
sampling=sampling,
u_range=u_range,
joint=False,
agent_id=i)
target_policy = DeterministicNNPolicy(env.env_specs,
hidden_layer_sizes=(M, M),
name='target_policy',
squash=squash,
squash_func=squash_func,
sampling=sampling,
u_range=u_range,
joint=False,
agent_id=i)
if k > 0:
policy, target_policy = get_level_k_policy(
env,
k,
M,
i,
u_range,
opponent_conditional_policy,
game_name=game_name)
joint_qf = NNJointQFunction(env_spec=env.env_specs,
hidden_layer_sizes=[M, M],
joint=joint,
agent_id=i)
target_joint_qf = NNJointQFunction(env_spec=env.env_specs,
hidden_layer_sizes=[M, M],
name='target_joint_qf',
joint=joint,
agent_id=i)
qf = NNQFunction(env_spec=env.env_specs,
hidden_layer_sizes=[M, M],
joint=False,
agent_id=i)
plotter = None
agent = MAVBAC(base_kwargs=base_kwargs,
agent_id=i,
env=env,
pool=pool,
joint_qf=joint_qf,
target_joint_qf=target_joint_qf,
qf=qf,
policy=policy,
target_policy=target_policy,
conditional_policy=opponent_conditional_policy,
plotter=plotter,
policy_lr=1e-2,
qf_lr=1e-2,
joint=False,
value_n_particles=16,
kernel_fn=adaptive_isotropic_gaussian_kernel,
kernel_n_particles=32,
kernel_update_ratio=0.5,
td_target_update_interval=1,
discount=0.95,
reward_scale=1,
tau=0.01,
save_full_state=False,
k=k,
aux=aux)
return agent