def main(c):
# Setup logz and save c
ps.configure_log(c)
# Create mdp and fix randomness
mdp = ps.setup_mdp(c['mdp'], c['seed'])
# Create learnable objects
ob_shape = mdp.ob_shape
ac_shape = mdp.ac_shape
if mdp.use_time_info:
ob_shape = (np.prod(ob_shape) + 1, )
# Define the learner
policy = RobustKerasMLPGassian(ob_shape,
ac_shape,
name='policy',
init_lstd=c['init_lstd'],
units=c['policy_units'])
vfn = SuperRobustKerasMLP(ob_shape, (1, ),
name='value function',
units=c['value_units'])
# Create algorithm
alg = PolicyGradient(policy,
vfn,
gamma=mdp.gamma,
horizon=mdp.horizon,
**c['algorithm'])
# Let's do some experiments!
exp = Exp.Experimenter(alg, mdp, c['experimenter']['ro_kwargs'])
exp.run(**c['experimenter']['run_kwargs'])
def main(c):
# Setup logz and save c
ps.configure_log(c)
# Create mdp and fix randomness
mdp = ps.setup_mdp(c['mdp'], c['seed'])
# Create learnable objects
ob_shape = mdp.ob_shape
ac_shape = mdp.ac_shape
if mdp.use_time_info:
ob_shape = (np.prod(ob_shape) + 1, )
# Define the learner
policy = RobustKerasMLPGassian(ob_shape,
ac_shape,
name='learner_policy',
init_lstd=c['init_lstd'],
units=c['policy_units'])
vfn = SuperRobustKerasMLP(ob_shape, (1, ),
name='learner_vfn',
units=c['value_units'])
policy, vfn = create_learner(c['mdp']['envid'], c['seed'], policy, vfn)
# Define experts
if c['use_experts']:
experts = create_experts(c['mdp']['envid'], **c['expert_info'])
if c['n_experts'] is not None and len(experts) > c['n_experts']:
experts = experts[:c['n_experts']]
if len(experts) < 1:
experts = None
else:
experts = None
# Create algorithm
ro_by_n_samples = c['experimenter']['ro_kwargs'] is not None
alg = Mamba(policy,
vfn,
experts=experts,
horizon=mdp.horizon,
gamma=mdp.gamma,
mix_unroll_kwargs={'ro_by_n_samples': ro_by_n_samples},
**c['algorithm'])
# Let's do some experiments!
exp = Exp.Experimenter(alg, mdp, c['experimenter']['ro_kwargs'])
exp.run(
seed=c['seed'],
**c['experimenter']['run_kwargs'],
)
def main(c):
# Setup logz and save c
ps.configure_log(c)
# Create mdp and fix randomness
mdp = ps.setup_mdp(c['mdp'], c['seed'])
# Create learnable objects
ob_shape = mdp.ob_shape
ac_shape = mdp.ac_shape
if mdp.use_time_info:
ob_shape = (np.prod(ob_shape)+1,)
# define expert
expert = RobustKerasMLPGassian(ob_shape, ac_shape, name='policy',
init_lstd=-1,
units=(64,))
expert.restore('./experts', name='cp1000_mlp_policy_64_seed_9')
expert.name = 'expert'
# define the learner
policy = RobustKerasMLPGassian(ob_shape, ac_shape, name='policy',
init_lstd=-1,
units=(128,128))
vfn = SuperRobustKerasMLP(ob_shape, (1,), name='expert value function',
units=(256,256))
# Create algorithm
alg = AggreVaTeD(policy, expert, vfn,
horizon=mdp.horizon, gamma=mdp.gamma,
**c['algorithm'])
# Let's do some experiments!
exp = Exp.Experimenter(alg, mdp, c['experimenter']['rollout_kwargs'])
exp.run(**c['experimenter']['run_kwargs'])