def create_policy(env, seed, c, name='learner_policy'):
pol_cls = getattr(Pol, c['policy_cls'])
ob_dim = env.observation_space.shape[0]
ac_dim = env.action_space.shape[0]
build_nor = Nor.create_build_nor_from_str(c['nor_cls'], c['nor_kwargs'])
policy = pol_cls(ob_dim,
ac_dim,
name=name,
seed=seed,
build_nor=build_nor,
**c['pol_kwargs'])
return policy
def create_advantage_estimator(policy,
adv_configs,
name='value_function_approximator'):
adv_configs = copy.deepcopy(adv_configs)
# create the value function SupervisedLearner
c = adv_configs['vfn_params']
build_nor = Nor.create_build_nor_from_str(c['nor_cls'], c['nor_kwargs'])
vfn_cls = getattr(Sup, c['fun_class'])
vfn = vfn_cls(policy.ob_dim,
1,
name=name,
build_nor=build_nor,
**c['fun_kwargs'])
# create the adv object
adv_configs.pop('vfn_params')
adv_configs['vfn'] = vfn
ae = AdvantageEstimator(policy, **adv_configs)
return ae
def create_predict(config, out=None):
# Create predict of a supervised learner.
# out_dim defaulted to be st_dim
st = env.reset()
st_dim, ac_dim = len(st), env.action_space.shape[0]
cls = getattr(Sup, config['fun_cls'])
build_nor = Nor.create_build_nor_from_str(config['nor_cls'], config['nor_kwargs'])
if out == 'dyn':
out_dim = st_dim
elif out == 'rw':
out_dim = 1
else:
raise ValueError('Invalid out: {}'.format(out))
# XXX learn residue for dynamics!
svl = cls(st_dim + ac_dim, out_dim, name=config['name'],
build_nor=build_nor, **config['fun_kwargs'])
return svl.predict
