pyRDDLGym-jax 2.0__py3-none-any.whl → 2.2__py3-none-any.whl

pyRDDLGym_jax/core/planner.py

@@ -47,7 +47,9 @@ import jax.random as random
 import numpy as np
 import optax
 import termcolor
-from tqdm import tqdm
+from tqdm import tqdm, TqdmWarning
+import warnings
+warnings.filterwarnings("ignore", category=TqdmWarning)
 
 from pyRDDLGym.core.compiler.model import RDDLPlanningModel, RDDLLiftedModel
 from pyRDDLGym.core.debug.logger import Logger
@@ -69,8 +71,7 @@ try:
     from pyRDDLGym_jax.core.visualization import JaxPlannerDashboard
 except Exception:
     raise_warning('Failed to load the dashboard visualization tool: '
-                  'please make sure you have installed the required packages.', 
-                  'red')
+                  'please make sure you have installed the required packages.', 'red')
     traceback.print_exc()
     JaxPlannerDashboard = None
 
@@ -133,7 +134,7 @@ def _load_config(config, args):
         comp_kwargs = model_args.get('complement_kwargs', {})
         compare_name = model_args.get('comparison', 'SigmoidComparison')
         compare_kwargs = model_args.get('comparison_kwargs', {})
-        sampling_name = model_args.get('sampling', 'GumbelSoftmax')
+        sampling_name = model_args.get('sampling', 'SoftRandomSampling')
         sampling_kwargs = model_args.get('sampling_kwargs', {})
         rounding_name = model_args.get('rounding', 'SoftRounding')
         rounding_kwargs = model_args.get('rounding_kwargs', {})
@@ -156,8 +157,7 @@ def _load_config(config, args):
         initializer = _getattr_any(
             packages=[initializers, hk.initializers], item=plan_initializer)
         if initializer is None:
-            raise_warning(
-                f'Ignoring invalid initializer <{plan_initializer}>.', 'red')
+            raise_warning(f'Ignoring invalid initializer <{plan_initializer}>.', 'red')
             del plan_kwargs['initializer']
         else:
             init_kwargs = plan_kwargs.pop('initializer_kwargs', {})
@@ -174,8 +174,7 @@ def _load_config(config, args):
         activation = _getattr_any(
             packages=[jax.nn, jax.numpy], item=plan_activation)
         if activation is None:
-            raise_warning(
-                f'Ignoring invalid activation <{plan_activation}>.', 'red')
+            raise_warning(f'Ignoring invalid activation <{plan_activation}>.', 'red')
             del plan_kwargs['activation']
         else:
             plan_kwargs['activation'] = activation
@@ -189,8 +188,7 @@ def _load_config(config, args):
     if planner_optimizer is not None:
         optimizer = _getattr_any(packages=[optax], item=planner_optimizer)
         if optimizer is None:
-            raise_warning(
-                f'Ignoring invalid optimizer <{planner_optimizer}>.', 'red')
+            raise_warning(f'Ignoring invalid optimizer <{planner_optimizer}>.', 'red')
             del planner_args['optimizer']
         else:
             planner_args['optimizer'] = optimizer
@@ -285,48 +283,14 @@ class JaxRDDLCompilerWithGrad(JaxRDDLCompiler):
         pvars_cast = set()
         for (var, values) in self.init_values.items():
             self.init_values[var] = np.asarray(values, dtype=self.REAL) 
-            if not np.issubdtype(np.atleast_1d(values).dtype, np.floating):
+            if not np.issubdtype(np.result_type(values), np.floating):
                 pvars_cast.add(var)
         if pvars_cast:
             raise_warning(f'JAX gradient compiler requires that initial values '
                           f'of p-variables {pvars_cast} be cast to float.')   
         
         # overwrite basic operations with fuzzy ones
-        self.RELATIONAL_OPS = {
-            '>=': logic.greater_equal,
-            '<=': logic.less_equal,
-            '<': logic.less,
-            '>': logic.greater,
-            '==': logic.equal,
-            '~=': logic.not_equal
-        }
-        self.LOGICAL_NOT = logic.logical_not
-        self.LOGICAL_OPS = {
-            '^': logic.logical_and,
-            '&': logic.logical_and,
-            '|': logic.logical_or,
-            '~': logic.xor,
-            '=>': logic.implies,
-            '<=>': logic.equiv
-        }
-        self.AGGREGATION_OPS['forall'] = logic.forall
-        self.AGGREGATION_OPS['exists'] = logic.exists
-        self.AGGREGATION_OPS['argmin'] = logic.argmin
-        self.AGGREGATION_OPS['argmax'] = logic.argmax
-        self.KNOWN_UNARY['sgn'] = logic.sgn
-        self.KNOWN_UNARY['floor'] = logic.floor
-        self.KNOWN_UNARY['ceil'] = logic.ceil
-        self.KNOWN_UNARY['round'] = logic.round
-        self.KNOWN_UNARY['sqrt'] = logic.sqrt
-        self.KNOWN_BINARY['div'] = logic.div
-        self.KNOWN_BINARY['mod'] = logic.mod
-        self.KNOWN_BINARY['fmod'] = logic.mod
-        self.IF_HELPER = logic.control_if
-        self.SWITCH_HELPER = logic.control_switch
-        self.BERNOULLI_HELPER = logic.bernoulli
-        self.DISCRETE_HELPER = logic.discrete
-        self.POISSON_HELPER = logic.poisson
-        self.GEOMETRIC_HELPER = logic.geometric
+        self.OPS = logic.get_operator_dicts()
         
     def _jax_stop_grad(self, jax_expr):        
         def _jax_wrapped_stop_grad(x, params, key):
@@ -575,7 +539,7 @@ class JaxStraightLinePlan(JaxPlan):
         def _jax_non_bool_param_to_action(var, param, hyperparams):
             if wrap_non_bool:
                 lower, upper = bounds_safe[var]
-                mb, ml, mu, mn = [mask.astype(compiled.REAL) 
+                mb, ml, mu, mn = [jnp.asarray(mask, dtype=compiled.REAL) 
                                   for mask in cond_lists[var]]       
                 action = (
                     mb * (lower + (upper - lower) * jax.nn.sigmoid(param)) + 
@@ -660,7 +624,7 @@ class JaxStraightLinePlan(JaxPlan):
                     action = _jax_non_bool_param_to_action(var, action, hyperparams)
                     action = jnp.clip(action, *bounds[var])
                     if ranges[var] == 'int':
-                        action = jnp.round(action).astype(compiled.INT)
+                        action = jnp.asarray(jnp.round(action), dtype=compiled.INT)
                     actions[var] = action
             return actions
         
@@ -961,12 +925,11 @@ class JaxDeepReactivePolicy(JaxPlan):
             non_bool_dims = 0
             for (var, values) in observed_vars.items():
                 if ranges[var] != 'bool':
-                    value_size = np.atleast_1d(values).size
+                    value_size = np.size(values)
                     if normalize_per_layer and value_size == 1:
                         raise_warning(
                             f'Cannot apply layer norm to state-fluent <{var}> '
-                            f'of size 1: setting normalize_per_layer = False.',
-                            'red')
+                            f'of size 1: setting normalize_per_layer = False.', 'red')
                         normalize_per_layer = False
                     non_bool_dims += value_size
             if not normalize_per_layer and non_bool_dims == 1:
@@ -990,9 +953,11 @@ class JaxDeepReactivePolicy(JaxPlan):
                     else:
                         if normalize and normalize_per_layer:
                             normalizer = hk.LayerNorm(
-                                axis=-1, param_axis=-1,
+                                axis=-1, 
+                                param_axis=-1,
                                 name=f'input_norm_{input_names[var]}',
-                                **self._normalizer_kwargs)
+                                **self._normalizer_kwargs
+                            )
                             state = normalizer(state)
                         states_non_bool.append(state)
                         non_bool_dims += state.size
@@ -1001,8 +966,11 @@ class JaxDeepReactivePolicy(JaxPlan):
             # optionally perform layer normalization on the non-bool inputs
             if normalize and not normalize_per_layer and non_bool_dims:
                 normalizer = hk.LayerNorm(
-                    axis=-1, param_axis=-1, name='input_norm',
-                    **self._normalizer_kwargs)
+                    axis=-1, 
+                    param_axis=-1, 
+                    name='input_norm',
+                    **self._normalizer_kwargs
+                )
                 normalized = normalizer(state[:non_bool_dims])
                 state = state.at[:non_bool_dims].set(normalized)
             return state
@@ -1021,7 +989,8 @@ class JaxDeepReactivePolicy(JaxPlan):
             actions = {}
             for (var, size) in layer_sizes.items():
                 linear = hk.Linear(size, name=layer_names[var], w_init=init)
-                reshape = hk.Reshape(output_shape=shapes[var], preserve_dims=-1,
+                reshape = hk.Reshape(output_shape=shapes[var], 
+                                     preserve_dims=-1,
                                      name=f'reshape_{layer_names[var]}')
                 output = reshape(linear(hidden))
                 if not shapes[var]:
@@ -1034,7 +1003,7 @@ class JaxDeepReactivePolicy(JaxPlan):
                 else:
                     if wrap_non_bool:
                         lower, upper = bounds_safe[var]
-                        mb, ml, mu, mn = [mask.astype(compiled.REAL) 
+                        mb, ml, mu, mn = [jnp.asarray(mask, dtype=compiled.REAL) 
                                           for mask in cond_lists[var]]       
                         action = (
                             mb * (lower + (upper - lower) * jax.nn.sigmoid(output)) + 
@@ -1048,8 +1017,7 @@ class JaxDeepReactivePolicy(JaxPlan):
             
             # for constraint satisfaction wrap bool actions with softmax
             if use_constraint_satisfaction:
-                linear = hk.Linear(
-                    bool_action_count, name='output_bool', w_init=init)
+                linear = hk.Linear(bool_action_count, name='output_bool', w_init=init)
                 output = jax.nn.softmax(linear(hidden))
                 actions[bool_key] = output
              
@@ -1087,8 +1055,7 @@ class JaxDeepReactivePolicy(JaxPlan):
         
         # test action prediction
         def _jax_wrapped_drp_predict_test(key, params, hyperparams, step, subs):
-            actions = _jax_wrapped_drp_predict_train(
-                key, params, hyperparams, step, subs)
+            actions = _jax_wrapped_drp_predict_train(key, params, hyperparams, step, subs)
             new_actions = {}
             for (var, action) in actions.items():
                 prange = ranges[var]
@@ -1096,7 +1063,7 @@ class JaxDeepReactivePolicy(JaxPlan):
                     new_action = action > 0.5
                 elif prange == 'int':
                     action = jnp.clip(action, *bounds[var])
-                    new_action = jnp.round(action).astype(compiled.INT)
+                    new_action = jnp.asarray(jnp.round(action), dtype=compiled.INT)
                 else:
                     new_action = jnp.clip(action, *bounds[var])
                 new_actions[var] = new_action
@@ -1247,17 +1214,22 @@ class GaussianPGPE(PGPE):
                  init_sigma: float=1.0,
                  sigma_range: Tuple[float, float]=(1e-5, 1e5),
                  scale_reward: bool=True,
+                 min_reward_scale: float=1e-5,
                  super_symmetric: bool=True,
                  super_symmetric_accurate: bool=True,
                  optimizer: Callable[..., optax.GradientTransformation]=optax.adam,
                  optimizer_kwargs_mu: Optional[Kwargs]=None,
-                 optimizer_kwargs_sigma: Optional[Kwargs]=None) -> None:
+                 optimizer_kwargs_sigma: Optional[Kwargs]=None,
+                 start_entropy_coeff: float=1e-3,
+                 end_entropy_coeff: float=1e-8,
+                 max_kl_update: Optional[float]=None) -> None:
         '''Creates a new Gaussian PGPE planner.
         
         :param batch_size: how many policy parameters to sample per optimization step
         :param init_sigma: initial standard deviation of Gaussian
         :param sigma_range: bounds to constrain standard deviation
         :param scale_reward: whether to apply reward scaling as in the paper
+        :param min_reward_scale: minimum reward scaling to avoid underflow
         :param super_symmetric: whether to use super-symmetric sampling as in the paper
         :param super_symmetric_accurate: whether to use the accurate formula for super-
         symmetric sampling or the simplified but biased formula
@@ -1266,6 +1238,9 @@ class GaussianPGPE(PGPE):
         factory for the mean optimizer
         :param optimizer_kwargs_sigma: a dictionary of parameters to pass to the SGD
         factory for the standard deviation optimizer
+        :param start_entropy_coeff: starting entropy regularization coeffient for Gaussian
+        :param end_entropy_coeff: ending entropy regularization coeffient for Gaussian
+        :param max_kl_update: bound on kl-divergence between parameter updates
         '''
         super().__init__()
 
@@ -1273,8 +1248,13 @@ class GaussianPGPE(PGPE):
         self.init_sigma = init_sigma
         self.sigma_range = sigma_range
         self.scale_reward = scale_reward
+        self.min_reward_scale = min_reward_scale
         self.super_symmetric = super_symmetric
         self.super_symmetric_accurate = super_symmetric_accurate
+
+        # entropy regularization penalty is decayed exponentially between these values
+        self.start_entropy_coeff = start_entropy_coeff
+        self.end_entropy_coeff = end_entropy_coeff
         
         # set optimizers
         if optimizer_kwargs_mu is None:
@@ -1284,36 +1264,62 @@ class GaussianPGPE(PGPE):
             optimizer_kwargs_sigma = {'learning_rate': 0.1}
         self.optimizer_kwargs_sigma = optimizer_kwargs_sigma
         self.optimizer_name = optimizer
-        mu_optimizer = optimizer(**optimizer_kwargs_mu)
-        sigma_optimizer = optimizer(**optimizer_kwargs_sigma)
+        try:
+            mu_optimizer = optax.inject_hyperparams(optimizer)(**optimizer_kwargs_mu)
+            sigma_optimizer = optax.inject_hyperparams(optimizer)(**optimizer_kwargs_sigma)
+        except Exception as _:
+            raise_warning(
+                f'Failed to inject hyperparameters into optax optimizer for PGPE, '
+                'rolling back to safer method: please note that kl-divergence '
+                'constraints will be disabled.', 'red')
+            mu_optimizer = optimizer(**optimizer_kwargs_mu)   
+            sigma_optimizer = optimizer(**optimizer_kwargs_sigma) 
+            max_kl_update = None
         self.optimizers = (mu_optimizer, sigma_optimizer)
+        self.max_kl = max_kl_update
     
     def __str__(self) -> str:
         return (f'PGPE hyper-parameters:\n'
-                f'    method         ={self.__class__.__name__}\n'
-                f'    batch_size     ={self.batch_size}\n'
-                f'    init_sigma     ={self.init_sigma}\n'
-                f'    sigma_range    ={self.sigma_range}\n'
-                f'    scale_reward   ={self.scale_reward}\n'
-                f'    super_symmetric={self.super_symmetric}\n'
-                f'        accurate   ={self.super_symmetric_accurate}\n'
-                f'    optimizer      ={self.optimizer_name}\n'
+                f'    method             ={self.__class__.__name__}\n'
+                f'    batch_size         ={self.batch_size}\n'
+                f'    init_sigma         ={self.init_sigma}\n'
+                f'    sigma_range        ={self.sigma_range}\n'
+                f'    scale_reward       ={self.scale_reward}\n'
+                f'    min_reward_scale   ={self.min_reward_scale}\n'
+                f'    super_symmetric    ={self.super_symmetric}\n'
+                f'        accurate       ={self.super_symmetric_accurate}\n'
+                f'    optimizer          ={self.optimizer_name}\n'
                 f'    optimizer_kwargs:\n'
                 f'        mu   ={self.optimizer_kwargs_mu}\n'
                 f'        sigma={self.optimizer_kwargs_sigma}\n'
+                f'    start_entropy_coeff={self.start_entropy_coeff}\n'
+                f'    end_entropy_coeff  ={self.end_entropy_coeff}\n'
+                f'    max_kl_update      ={self.max_kl}\n'
         )
 
     def compile(self, loss_fn: Callable, projection: Callable, real_dtype: Type) -> None:
-        MIN_NORM = 1e-5
         sigma0 = self.init_sigma
         sigma_range = self.sigma_range
         scale_reward = self.scale_reward
+        min_reward_scale = self.min_reward_scale
         super_symmetric = self.super_symmetric
         super_symmetric_accurate = self.super_symmetric_accurate
         batch_size = self.batch_size
         optimizers = (mu_optimizer, sigma_optimizer) = self.optimizers
-
-        # initializer
+        max_kl = self.max_kl
+        
+        # entropy regularization penalty is decayed exponentially by elapsed budget
+        start_entropy_coeff = self.start_entropy_coeff
+        if start_entropy_coeff == 0:
+            entropy_coeff_decay = 0
+        else:
+            entropy_coeff_decay = (self.end_entropy_coeff / start_entropy_coeff) ** 0.01
+        
+        # ***********************************************************************
+        # INITIALIZATION OF POLICY
+        #
+        # ***********************************************************************
+        
         def _jax_wrapped_pgpe_init(key, policy_params):
             mu = policy_params
             sigma = jax.tree_map(lambda x: sigma0 * jnp.ones_like(x), mu)
@@ -1324,7 +1330,11 @@ class GaussianPGPE(PGPE):
         
         self._initializer = jax.jit(_jax_wrapped_pgpe_init)
 
-        # parameter sampling functions
+        # ***********************************************************************
+        # PARAMETER SAMPLING FUNCTIONS
+        #
+        # ***********************************************************************
+
         def _jax_wrapped_mu_noise(key, sigma):
             return sigma * random.normal(key, shape=jnp.shape(sigma), dtype=real_dtype)
 
@@ -1334,19 +1344,20 @@ class GaussianPGPE(PGPE):
             a = (sigma - jnp.abs(epsilon)) / sigma
             if super_symmetric_accurate:
                 aa = jnp.abs(a)
+                aa3 = jnp.power(aa, 3)
                 epsilon_star = jnp.sign(epsilon) * phi * jnp.where(
                     a <= 0,
-                    jnp.exp(c1 * aa * (aa * aa - 1) / jnp.log(aa + 1e-10) + c2 * aa),
-                    jnp.exp(aa - c3 * aa * jnp.log(1.0 - jnp.power(aa, 3) + 1e-10))
+                    jnp.exp(c1 * (aa3 - aa) / jnp.log(aa + 1e-10) + c2 * aa),
+                    jnp.exp(aa - c3 * aa * jnp.log(1.0 - aa3 + 1e-10))
                 )
             else:
                 epsilon_star = jnp.sign(epsilon) * phi * jnp.exp(a)
             return epsilon_star
 
         def _jax_wrapped_sample_params(key, mu, sigma):
-            keys = random.split(key, num=len(jax.tree_util.tree_leaves(mu)))
-            keys_pytree = jax.tree_util.tree_unflatten(
-                treedef=jax.tree_util.tree_structure(mu), leaves=keys)
+            treedef = jax.tree_util.tree_structure(sigma)
+            keys = random.split(key, num=treedef.num_leaves)
+            keys_pytree = jax.tree_util.tree_unflatten(treedef=treedef, leaves=keys)
             epsilon = jax.tree_map(_jax_wrapped_mu_noise, keys_pytree, sigma)
             p1 = jax.tree_map(jnp.add, mu, epsilon)
             p2 = jax.tree_map(jnp.subtract, mu, epsilon)
@@ -1356,14 +1367,18 @@ class GaussianPGPE(PGPE):
                 p4 = jax.tree_map(jnp.subtract, mu, epsilon_star)
             else:
                 epsilon_star, p3, p4 = epsilon, p1, p2
-            return (p1, p2, p3, p4), (epsilon, epsilon_star)
+            return p1, p2, p3, p4, epsilon, epsilon_star
                         
-        # policy gradient update functions
+        # ***********************************************************************
+        # POLICY GRADIENT CALCULATION
+        #
+        # ***********************************************************************
+
         def _jax_wrapped_mu_grad(epsilon, epsilon_star, r1, r2, r3, r4, m):
             if super_symmetric:
                 if scale_reward:
-                    scale1 = jnp.maximum(MIN_NORM, m - (r1 + r2) / 2)
-                    scale2 = jnp.maximum(MIN_NORM, m - (r3 + r4) / 2)
+                    scale1 = jnp.maximum(min_reward_scale, m - (r1 + r2) / 2)
+                    scale2 = jnp.maximum(min_reward_scale, m - (r3 + r4) / 2)
                 else:
                     scale1 = scale2 = 1.0
                 r_mu1 = (r1 - r2) / (2 * scale1)
@@ -1371,37 +1386,37 @@ class GaussianPGPE(PGPE):
                 grad = -(r_mu1 * epsilon + r_mu2 * epsilon_star)
             else:
                 if scale_reward:
-                    scale = jnp.maximum(MIN_NORM, m - (r1 + r2) / 2)
+                    scale = jnp.maximum(min_reward_scale, m - (r1 + r2) / 2)
                 else:
                     scale = 1.0
                 r_mu = (r1 - r2) / (2 * scale)
                 grad = -r_mu * epsilon
             return grad
         
-        def _jax_wrapped_sigma_grad(epsilon, epsilon_star, sigma, r1, r2, r3, r4, m):
+        def _jax_wrapped_sigma_grad(epsilon, epsilon_star, sigma, r1, r2, r3, r4, m, ent):
             if super_symmetric:
                 mask = r1 + r2 >= r3 + r4
                 epsilon_tau = mask * epsilon + (1 - mask) * epsilon_star
-                s = epsilon_tau * epsilon_tau / sigma - sigma
+                s = jnp.square(epsilon_tau) / sigma - sigma
                 if scale_reward:
-                    scale = jnp.maximum(MIN_NORM, m - (r1 + r2 + r3 + r4) / 4)
+                    scale = jnp.maximum(min_reward_scale, m - (r1 + r2 + r3 + r4) / 4)
                 else:
                     scale = 1.0
                 r_sigma = ((r1 + r2) - (r3 + r4)) / (4 * scale)
             else:
-                s = epsilon * epsilon / sigma - sigma
+                s = jnp.square(epsilon) / sigma - sigma
                 if scale_reward:
-                    scale = jnp.maximum(MIN_NORM, jnp.abs(m))
+                    scale = jnp.maximum(min_reward_scale, jnp.abs(m))
                 else:
                     scale = 1.0
                 r_sigma = (r1 + r2) / (2 * scale)
-            grad = -r_sigma * s
+            grad = -(r_sigma * s + ent / sigma)
             return grad
             
-        def _jax_wrapped_pgpe_grad(key, mu, sigma, r_max, 
+        def _jax_wrapped_pgpe_grad(key, mu, sigma, r_max, ent,
                                    policy_hyperparams, subs, model_params):
             key, subkey = random.split(key)
-            (p1, p2, p3, p4), (epsilon, epsilon_star) = _jax_wrapped_sample_params(
+            p1, p2, p3, p4, epsilon, epsilon_star = _jax_wrapped_sample_params(
                 key, mu, sigma)
             r1 = -loss_fn(subkey, p1, policy_hyperparams, subs, model_params)[0]
             r2 = -loss_fn(subkey, p2, policy_hyperparams, subs, model_params)[0]
@@ -1419,42 +1434,76 @@ class GaussianPGPE(PGPE):
                 epsilon, epsilon_star
             ) 
             grad_sigma = jax.tree_map(
-                partial(_jax_wrapped_sigma_grad, r1=r1, r2=r2, r3=r3, r4=r4, m=r_max), 
+                partial(_jax_wrapped_sigma_grad, 
+                        r1=r1, r2=r2, r3=r3, r4=r4, m=r_max, ent=ent), 
                 epsilon, epsilon_star, sigma
             )
             return grad_mu, grad_sigma, r_max
 
-        def _jax_wrapped_pgpe_grad_batched(key, pgpe_params, r_max, 
+        def _jax_wrapped_pgpe_grad_batched(key, pgpe_params, r_max, ent,
                                            policy_hyperparams, subs, model_params):
             mu, sigma = pgpe_params
             if batch_size == 1:
                 mu_grad, sigma_grad, new_r_max = _jax_wrapped_pgpe_grad(
-                    key, mu, sigma, r_max, policy_hyperparams, subs, model_params)
+                    key, mu, sigma, r_max, ent, policy_hyperparams, subs, model_params)
             else:
                 keys = random.split(key, num=batch_size)
                 mu_grads, sigma_grads, r_maxs = jax.vmap(
                     _jax_wrapped_pgpe_grad, 
-                    in_axes=(0, None, None, None, None, None, None)
-                )(keys, mu, sigma, r_max, policy_hyperparams, subs, model_params)
-                mu_grad = jax.tree_map(partial(jnp.mean, axis=0), mu_grads)
-                sigma_grad = jax.tree_map(partial(jnp.mean, axis=0), sigma_grads)
+                    in_axes=(0, None, None, None, None, None, None, None)
+                )(keys, mu, sigma, r_max, ent, policy_hyperparams, subs, model_params)
+                mu_grad, sigma_grad = jax.tree_map(
+                    partial(jnp.mean, axis=0), (mu_grads, sigma_grads))
                 new_r_max = jnp.max(r_maxs)
             return mu_grad, sigma_grad, new_r_max
+        
+        # ***********************************************************************
+        # PARAMETER UPDATE
+        #
+        # ***********************************************************************
 
-        def _jax_wrapped_pgpe_update(key, pgpe_params, r_max, 
+        def _jax_wrapped_pgpe_kl_term(mu, sigma, old_mu, old_sigma):
+            return 0.5 * jnp.sum(2 * jnp.log(sigma / old_sigma) + 
+                                 jnp.square(old_sigma / sigma) + 
+                                 jnp.square((mu - old_mu) / sigma) - 1)
+        
+        def _jax_wrapped_pgpe_update(key, pgpe_params, r_max, progress,
                                      policy_hyperparams, subs, model_params, 
                                      pgpe_opt_state):
+            # regular update
             mu, sigma = pgpe_params
             mu_state, sigma_state = pgpe_opt_state
+            ent = start_entropy_coeff * jnp.power(entropy_coeff_decay, progress)
             mu_grad, sigma_grad, new_r_max = _jax_wrapped_pgpe_grad_batched(
-                key, pgpe_params, r_max, policy_hyperparams, subs, model_params)
+                key, pgpe_params, r_max, ent, policy_hyperparams, subs, model_params)
             mu_updates, new_mu_state = mu_optimizer.update(mu_grad, mu_state, params=mu) 
             sigma_updates, new_sigma_state = sigma_optimizer.update(
                 sigma_grad, sigma_state, params=sigma) 
             new_mu = optax.apply_updates(mu, mu_updates)
-            new_mu, converged = projection(new_mu, policy_hyperparams)
             new_sigma = optax.apply_updates(sigma, sigma_updates)
             new_sigma = jax.tree_map(lambda x: jnp.clip(x, *sigma_range), new_sigma)
+            
+            # respect KL divergence contraint with old parameters
+            if max_kl is not None:
+                old_mu_lr = new_mu_state.hyperparams['learning_rate']
+                old_sigma_lr = new_sigma_state.hyperparams['learning_rate']
+                kl_terms = jax.tree_map(
+                    _jax_wrapped_pgpe_kl_term, new_mu, new_sigma, mu, sigma)
+                total_kl = jax.tree_util.tree_reduce(jnp.add, kl_terms)
+                kl_reduction = jnp.minimum(1.0, jnp.sqrt(max_kl / total_kl))
+                mu_state.hyperparams['learning_rate'] = old_mu_lr * kl_reduction
+                sigma_state.hyperparams['learning_rate'] = old_sigma_lr * kl_reduction
+                mu_updates, new_mu_state = mu_optimizer.update(mu_grad, mu_state, params=mu) 
+                sigma_updates, new_sigma_state = sigma_optimizer.update(
+                    sigma_grad, sigma_state, params=sigma) 
+                new_mu = optax.apply_updates(mu, mu_updates)
+                new_sigma = optax.apply_updates(sigma, sigma_updates)
+                new_sigma = jax.tree_map(lambda x: jnp.clip(x, *sigma_range), new_sigma)
+                new_mu_state.hyperparams['learning_rate'] = old_mu_lr
+                new_sigma_state.hyperparams['learning_rate'] = old_sigma_lr
+
+            # apply projection step and finalize results
+            new_mu, converged = projection(new_mu, policy_hyperparams)
             new_pgpe_params = (new_mu, new_sigma)
             new_pgpe_opt_state = (new_mu_state, new_sigma_state)
             policy_params = new_mu
@@ -1463,6 +1512,71 @@ class GaussianPGPE(PGPE):
         self._update = jax.jit(_jax_wrapped_pgpe_update)
 
 
+# ***********************************************************************
+# ALL VERSIONS OF RISK FUNCTIONS
+# 
+# Based on the original paper "A Distributional Framework for Risk-Sensitive 
+# End-to-End Planning in Continuous MDPs" by Patton et al., AAAI 2022.
+#
+# Original risk functions:
+# - entropic utility
+# - mean-variance
+# - mean-semideviation
+# - conditional value at risk with straight-through gradient trick
+#
+# ***********************************************************************
+
+
+@jax.jit
+def entropic_utility(returns: jnp.ndarray, beta: float) -> float:
+    return (-1.0 / beta) * jax.scipy.special.logsumexp(
+        -beta * returns, b=1.0 / returns.size)
+
+
+@jax.jit
+def mean_variance_utility(returns: jnp.ndarray, beta: float) -> float:
+    return jnp.mean(returns) - 0.5 * beta * jnp.var(returns)
+
+
+@jax.jit
+def mean_deviation_utility(returns: jnp.ndarray, beta: float) -> float:
+    return jnp.mean(returns) - 0.5 * beta * jnp.std(returns)
+
+
+@jax.jit
+def mean_semideviation_utility(returns: jnp.ndarray, beta: float) -> float:
+    mu = jnp.mean(returns)
+    msd = jnp.sqrt(jnp.mean(jnp.square(jnp.minimum(0.0, returns - mu))))
+    return mu - 0.5 * beta * msd
+
+
+@jax.jit
+def mean_semivariance_utility(returns: jnp.ndarray, beta: float) -> float:
+    mu = jnp.mean(returns)
+    msv = jnp.mean(jnp.square(jnp.minimum(0.0, returns - mu)))
+    return mu - 0.5 * beta * msv
+
+
+@jax.jit
+def cvar_utility(returns: jnp.ndarray, alpha: float) -> float:
+    var = jnp.percentile(returns, q=100 * alpha)
+    mask = returns <= var
+    weights = mask / jnp.maximum(1, jnp.sum(mask))
+    return jnp.sum(returns * weights)
+
+
+UTILITY_LOOKUP = {
+    'mean': jnp.mean,
+    'mean_var': mean_variance_utility,
+    'mean_std': mean_deviation_utility,
+    'mean_semivar': mean_semivariance_utility,
+    'mean_semidev': mean_semideviation_utility,
+    'entropic': entropic_utility,
+    'exponential': entropic_utility,
+    'cvar': cvar_utility
+}
+
+
 # ***********************************************************************
 # ALL VERSIONS OF JAX PLANNER
 # 
@@ -1525,8 +1639,7 @@ class JaxBackpropPlanner:
         reward as a form of normalization
         :param utility: how to aggregate return observations to compute utility
         of a policy or plan; must be either a function mapping jax array to a 
-        scalar, or a a string identifying the utility function by name 
-        ("mean", "mean_var", "entropic", or "cvar" are currently supported)
+        scalar, or a a string identifying the utility function by name
         :param utility_kwargs: additional keyword arguments to pass hyper-
         parameters to the utility function call
         :param cpfs_without_grad: which CPFs do not have gradients (use straight
@@ -1584,18 +1697,11 @@ class JaxBackpropPlanner:
         # set utility
         if isinstance(utility, str):
             utility = utility.lower()
-            if utility == 'mean':
-                utility_fn = jnp.mean
-            elif utility == 'mean_var':
-                utility_fn = mean_variance_utility
-            elif utility == 'entropic':
-                utility_fn = entropic_utility
-            elif utility == 'cvar':
-                utility_fn = cvar_utility
-            else:
+            utility_fn = UTILITY_LOOKUP.get(utility, None)
+            if utility_fn is None:
                 raise RDDLNotImplementedError(
-                    f'Utility function <{utility}> is not supported: '
-                    'must be one of ["mean", "mean_var", "entropic", "cvar"].')
+                    f'Utility <{utility}> is not supported, '
+                    f'must be one of {list(UTILITY_LOOKUP.keys())}.')
         else:
             utility_fn = utility
         self.utility = utility_fn
@@ -1746,7 +1852,6 @@ r"""
         
         # optimization
         self.update = self._jax_update(train_loss)
-        self.check_zero_grad = self._jax_check_zero_gradients()
 
         # pgpe option
         if self.use_pgpe:
@@ -1809,6 +1914,12 @@ r"""
         projection = self.plan.projection
         use_ls = self.line_search_kwargs is not None
         
+        # check if the gradients are all zeros
+        def _jax_wrapped_zero_gradients(grad):
+            leaves, _ = jax.tree_util.tree_flatten(
+                jax.tree_map(lambda g: jnp.allclose(g, 0), grad))
+            return jnp.all(jnp.asarray(leaves))
+        
         # calculate the plan gradient w.r.t. return loss and update optimizer
         # also perform a projection step to satisfy constraints on actions
         def _jax_wrapped_loss_swapped(policy_params, key, policy_hyperparams,
@@ -1833,23 +1944,12 @@ r"""
             policy_params, converged = projection(policy_params, policy_hyperparams)
             log['grad'] = grad
             log['updates'] = updates
+            zero_grads = _jax_wrapped_zero_gradients(grad)
             return policy_params, converged, opt_state, opt_aux, \
-                loss_val, log, model_params
+                loss_val, log, model_params, zero_grads
         
         return jax.jit(_jax_wrapped_plan_update)
             
-    def _jax_check_zero_gradients(self):
-        
-        def _jax_wrapped_zero_gradient(grad):
-            return jnp.allclose(grad, 0)
-
-        def _jax_wrapped_zero_gradients(grad):
-            leaves, _ = jax.tree_util.tree_flatten(
-                jax.tree_map(_jax_wrapped_zero_gradient, grad))
-            return jnp.all(jnp.asarray(leaves))
-        
-        return jax.jit(_jax_wrapped_zero_gradients)
-
     def _batched_init_subs(self, subs): 
         rddl = self.rddl
         n_train, n_test = self.batch_size_train, self.batch_size_test
@@ -1865,7 +1965,7 @@ r"""
                     f'{set(self.test_compiled.init_values.keys())}.')
             value = np.reshape(value, newshape=np.shape(init_value))[np.newaxis, ...]
             train_value = np.repeat(value, repeats=n_train, axis=0)
-            train_value = train_value.astype(self.compiled.REAL)
+            train_value = np.asarray(train_value, dtype=self.compiled.REAL)
             init_train[name] = train_value
             init_test[name] = np.repeat(value, repeats=n_test, axis=0)
         
@@ -2153,11 +2253,12 @@ r"""
         # ======================================================================
         
         # initialize running statistics
-        best_params, best_loss, best_grad = policy_params, jnp.inf, jnp.inf
+        best_params, best_loss, best_grad = policy_params, jnp.inf, None
         last_iter_improve = 0
         rolling_test_loss = RollingMean(test_rolling_window)
         log = {}
         status = JaxPlannerStatus.NORMAL
+        progress_percent = 0
         
         # initialize stopping criterion
         if stopping_rule is not None:
@@ -2169,16 +2270,19 @@ r"""
                 dashboard_id, dashboard.get_planner_info(self), 
                 key=dash_key, viz=self.dashboard_viz)
         
+        # progress bar
+        if print_progress:
+            progress_bar = tqdm(None, total=100, position=tqdm_position, 
+                                bar_format='{l_bar}{bar}| {elapsed} {postfix}')
+        else:
+            progress_bar = None
+        position_str = '' if tqdm_position is None else f'[{tqdm_position}]'
+        
         # ======================================================================
         # MAIN TRAINING LOOP BEGINS
         # ======================================================================
         
-        iters = range(epochs)
-        if print_progress:
-            iters = tqdm(iters, total=100, position=tqdm_position)
-        position_str = '' if tqdm_position is None else f'[{tqdm_position}]'
-        
-        for it in iters:
+        for it in range(epochs):
             
             # ==================================================================
             # NEXT GRADIENT DESCENT STEP
@@ -2189,8 +2293,9 @@ r"""
             # update the parameters of the plan
             key, subkey = random.split(key)
             (policy_params, converged, opt_state, opt_aux, train_loss, train_log, 
-             model_params) = self.update(subkey, policy_params, policy_hyperparams,
-                                         train_subs, model_params, opt_state, opt_aux)
+             model_params, zero_grads) = self.update(
+                 subkey, policy_params, policy_hyperparams, train_subs, model_params, 
+                 opt_state, opt_aux)
             test_loss, (test_log, model_params_test) = self.test_loss(
                 subkey, policy_params, policy_hyperparams, test_subs, model_params_test)
             test_loss_smooth = rolling_test_loss.update(test_loss)
@@ -2200,8 +2305,9 @@ r"""
             if self.use_pgpe:
                 key, subkey = random.split(key)
                 pgpe_params, r_max, pgpe_opt_state, pgpe_param, pgpe_converged = \
-                    self.pgpe.update(subkey, pgpe_params, r_max, policy_hyperparams, 
-                                     test_subs, model_params, pgpe_opt_state)
+                    self.pgpe.update(subkey, pgpe_params, r_max, progress_percent, 
+                                     policy_hyperparams, test_subs, model_params_test, 
+                                     pgpe_opt_state)
                 pgpe_loss, _ = self.test_loss(
                     subkey, pgpe_param, policy_hyperparams, test_subs, model_params_test)
                 pgpe_loss_smooth = rolling_pgpe_loss.update(pgpe_loss)
@@ -2228,7 +2334,7 @@ r"""
             # ==================================================================
             
             # no progress
-            if (not pgpe_improve) and self.check_zero_grad(train_log['grad']):
+            if (not pgpe_improve) and zero_grads:
                 status = JaxPlannerStatus.NO_PROGRESS
              
             # constraint satisfaction problem
@@ -2256,7 +2362,8 @@ r"""
                 status = JaxPlannerStatus.ITER_BUDGET_REACHED
             
             # build a callback
-            progress_percent = int(100 * min(1, max(elapsed / train_seconds, it / epochs)))
+            progress_percent = 100 * min(
+                1, max(0, elapsed / train_seconds, it / (epochs - 1)))
             callback = {
                 'status': status,
                 'iteration': it,
@@ -2279,19 +2386,22 @@ r"""
                 'train_log': train_log,
                 **test_log
             }
-            
+
             # stopping condition reached
             if stopping_rule is not None and stopping_rule.monitor(callback):
                 callback['status'] = status = JaxPlannerStatus.STOPPING_RULE_REACHED  
             
             # if the progress bar is used
             if print_progress:
-                iters.n = progress_percent
-                iters.set_description(
+                progress_bar.set_description(
                     f'{position_str} {it:6} it / {-train_loss:14.5f} train / '
                     f'{-test_loss_smooth:14.5f} test / {-best_loss:14.5f} best / '
-                    f'{status.value} status / {total_pgpe_it:6} pgpe'
+                    f'{status.value} status / {total_pgpe_it:6} pgpe',
+                    refresh=False
                 )
+                progress_bar.set_postfix_str(
+                    f"{(it + 1) / (elapsed + 1e-6):.2f}it/s", refresh=False)
+                progress_bar.update(progress_percent - progress_bar.n)
             
             # dash-board
             if dashboard is not None:
@@ -2312,7 +2422,7 @@ r"""
 
         # release resources
         if print_progress:
-            iters.close()
+            progress_bar.close()
         
         # validate the test return
         if log:
@@ -2332,7 +2442,7 @@ r"""
                 last_iter_improve, -train_loss, -test_loss_smooth, -best_loss, grad_norm)
             print(f'summary of optimization:\n'
                   f'    status        ={status}\n'
-                  f'    time          ={elapsed:.6f} sec.\n'
+                  f'    time          ={elapsed:.3f} sec.\n'
                   f'    iterations    ={it}\n'
                   f'    best objective={-best_loss:.6f}\n'
                   f'    best grad norm={grad_norm}\n'
@@ -2358,12 +2468,12 @@ r"""
                 return termcolor.colored(
                     '[FAILURE] no progress was made '
                     f'and max grad norm {max_grad_norm:.6f} was zero: '
-                    'the solver was likely stuck in a plateau.', 'red')
+                    'solver likely stuck in a plateau.', 'red')
             else:
                 return termcolor.colored(
                     '[FAILURE] no progress was made '
                     f'but max grad norm {max_grad_norm:.6f} was non-zero: '
-                    'the learning rate or other hyper-parameters were likely suboptimal.', 
+                    'learning rate or other hyper-parameters likely suboptimal.', 
                     'red')
         
         # model is likely poor IF:
@@ -2372,8 +2482,8 @@ r"""
             return termcolor.colored(
                 '[WARNING] progress was made '
                 f'but relative train-test error {validation_error:.6f} was high: '
-                'model relaxation around the solution was poor '
-                'or the batch size was too small.', 'yellow')
+                'poor model relaxation around solution or batch size too small.', 
+                'yellow')
         
         # model likely did not converge IF:
         # 1. the max grad relative to the return is high
@@ -2383,9 +2493,9 @@ r"""
                 return termcolor.colored(
                     '[WARNING] progress was made '
                     f'but max grad norm {max_grad_norm:.6f} was high: '
-                    'the solution was likely locally suboptimal, '
-                    'or the relaxed model was not smooth around the solution, '
-                    'or the batch size was too small.', 'yellow')
+                    'solution locally suboptimal '
+                    'or relaxed model not smooth around solution '
+                    'or batch size too small.', 'yellow')
         
         # likely successful
         return termcolor.colored(
@@ -2412,8 +2522,7 @@ r"""
         for (var, values) in subs.items():
             
             # must not be grounded
-            if RDDLPlanningModel.FLUENT_SEP in var \
-            or RDDLPlanningModel.OBJECT_SEP in var:
+            if RDDLPlanningModel.FLUENT_SEP in var or RDDLPlanningModel.OBJECT_SEP in var:
                 raise ValueError(f'State dictionary passed to the JAX policy is '
                                  f'grounded, since it contains the key <{var}>, '
                                  f'but a vectorized environment is required: '
@@ -2421,9 +2530,8 @@ r"""
             
             # must be numeric array
             # exception is for POMDPs at 1st epoch when observ-fluents are None
-            dtype = np.atleast_1d(values).dtype
-            if not np.issubdtype(dtype, np.number) \
-            and not np.issubdtype(dtype, np.bool_):
+            dtype = np.result_type(values)
+            if not np.issubdtype(dtype, np.number) and not np.issubdtype(dtype, np.bool_):
                 if step == 0 and var in self.rddl.observ_fluents:
                     subs[var] = self.test_compiled.init_values[var]
                 else:
@@ -2435,40 +2543,7 @@ r"""
         actions = self.test_policy(key, params, policy_hyperparams, step, subs)
         actions = jax.tree_map(np.asarray, actions)
         return actions      
-    
-
-# ***********************************************************************
-# ALL VERSIONS OF RISK FUNCTIONS
-# 
-# Based on the original paper "A Distributional Framework for Risk-Sensitive 
-# End-to-End Planning in Continuous MDPs" by Patton et al., AAAI 2022.
-#
-# Original risk functions:
-# - entropic utility
-# - mean-variance approximation
-# - conditional value at risk with straight-through gradient trick
-#
-# ***********************************************************************
-
-
-@jax.jit
-def entropic_utility(returns: jnp.ndarray, beta: float) -> float:
-    return (-1.0 / beta) * jax.scipy.special.logsumexp(
-        -beta * returns, b=1.0 / returns.size)
-
-
-@jax.jit
-def mean_variance_utility(returns: jnp.ndarray, beta: float) -> float:
-    return jnp.mean(returns) - 0.5 * beta * jnp.var(returns)
-    
-
-@jax.jit
-def cvar_utility(returns: jnp.ndarray, alpha: float) -> float:
-    var = jnp.percentile(returns, q=100 * alpha)
-    mask = returns <= var
-    weights = mask / jnp.maximum(1, jnp.sum(mask))
-    return jnp.sum(returns * weights)
-   
+       
 
 # ***********************************************************************
 # ALL VERSIONS OF CONTROLLERS
@@ -2580,8 +2655,7 @@ class JaxOnlineController(BaseAgent):
         self.callback = callback
         params = callback['best_params']
         self.key, subkey = random.split(self.key)
-        actions = planner.get_action(
-            subkey, params, 0, state, self.eval_hyperparams)
+        actions = planner.get_action(subkey, params, 0, state, self.eval_hyperparams)
         if self.warm_start:
             self.guess = planner.plan.guess_next_epoch(params)
         return actions