pyRDDLGym-jax 2.1__tar.gz → 2.2__tar.gz

{pyrddlgym_jax-2.1 → pyrddlgym_jax-2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: pyRDDLGym-jax
-Version: 2.1
+Version: 2.2
 Summary: pyRDDLGym-jax: automatic differentiation for solving sequential planning problems in JAX.
 Home-page: https://github.com/pyrddlgym-project/pyRDDLGym-jax
 Author: Michael Gimelfarb, Ayal Taitler, Scott Sanner
@@ -58,8 +58,11 @@ Dynamic: summary
 
 Purpose:
 
-1. automatic translation of any RDDL description file into a differentiable simulator in JAX
-2. flexible policy class representations, automatic model relaxations for working in discrete and hybrid domains, and Bayesian hyper-parameter tuning.
+1. automatic translation of RDDL description files into differentiable JAX simulators
+2. implementation of (highly configurable) operator relaxations for working in discrete and hybrid domains
+3. flexible policy representations and automated Bayesian hyper-parameter tuning
+4. interactive dashboard for dyanmic visualization and debugging
+5. hybridization with parameter-exploring policy gradients.
 
 Some demos of solved problems by JaxPlan:
 
@@ -235,8 +238,23 @@ More documentation about this and other new features will be coming soon.
 
 ## Tuning the Planner
 
-It is easy to tune the planner's hyper-parameters efficiently and automatically using Bayesian optimization.
-To do this, first create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
+A basic run script is provided to run automatic Bayesian hyper-parameter tuning for the most sensitive parameters of JaxPlan:
+
+```shell
+jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
+```
+
+where:
+- ``domain`` is the domain identifier as specified in rddlrepository
+- ``instance`` is the instance identifier
+- ``method`` is the planning method to use (i.e. drp, slp, replan)
+- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
+- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
+- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
+- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
+
+It is easy to tune a custom range of the planner's hyper-parameters efficiently. 
+First create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
 
 ```ini
 [Model]
@@ -260,7 +278,7 @@ train_on_reset=True
 
 would allow to tune the sharpness of model relaxations, and the learning rate of the optimizer.
 
-Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand:
+Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand, and run the optimizer:
 
 ```python
 import pyRDDLGym
@@ -292,22 +310,7 @@ tuning = JaxParameterTuning(env=env,
                             gp_iters=iters)
 tuning.tune(key=42, log_file='path/to/log.csv')
 ```
- 
-A basic run script is provided to run the automatic hyper-parameter tuning for the most sensitive parameters of JaxPlan:
-
-```shell
-jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
-```
-
-where:
-- ``domain`` is the domain identifier as specified in rddlrepository
-- ``instance`` is the instance identifier
-- ``method`` is the planning method to use (i.e. drp, slp, replan)
-- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
-- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
-- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
-- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
- 
+  
 
 ## Simulation
 

{pyrddlgym_jax-2.1 → pyrddlgym_jax-2.2}/README.md

@@ -12,8 +12,11 @@
 
 Purpose:
 
-1. automatic translation of any RDDL description file into a differentiable simulator in JAX
-2. flexible policy class representations, automatic model relaxations for working in discrete and hybrid domains, and Bayesian hyper-parameter tuning.
+1. automatic translation of RDDL description files into differentiable JAX simulators
+2. implementation of (highly configurable) operator relaxations for working in discrete and hybrid domains
+3. flexible policy representations and automated Bayesian hyper-parameter tuning
+4. interactive dashboard for dyanmic visualization and debugging
+5. hybridization with parameter-exploring policy gradients.
 
 Some demos of solved problems by JaxPlan:
 
@@ -189,8 +192,23 @@ More documentation about this and other new features will be coming soon.
 
 ## Tuning the Planner
 
-It is easy to tune the planner's hyper-parameters efficiently and automatically using Bayesian optimization.
-To do this, first create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
+A basic run script is provided to run automatic Bayesian hyper-parameter tuning for the most sensitive parameters of JaxPlan:
+
+```shell
+jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
+```
+
+where:
+- ``domain`` is the domain identifier as specified in rddlrepository
+- ``instance`` is the instance identifier
+- ``method`` is the planning method to use (i.e. drp, slp, replan)
+- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
+- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
+- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
+- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
+
+It is easy to tune a custom range of the planner's hyper-parameters efficiently. 
+First create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
 
 ```ini
 [Model]
@@ -214,7 +232,7 @@ train_on_reset=True
 
 would allow to tune the sharpness of model relaxations, and the learning rate of the optimizer.
 
-Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand:
+Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand, and run the optimizer:
 
 ```python
 import pyRDDLGym
@@ -246,22 +264,7 @@ tuning = JaxParameterTuning(env=env,
                             gp_iters=iters)
 tuning.tune(key=42, log_file='path/to/log.csv')
 ```
- 
-A basic run script is provided to run the automatic hyper-parameter tuning for the most sensitive parameters of JaxPlan:
-
-```shell
-jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
-```
-
-where:
-- ``domain`` is the domain identifier as specified in rddlrepository
-- ``instance`` is the instance identifier
-- ``method`` is the planning method to use (i.e. drp, slp, replan)
-- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
-- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
-- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
-- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
- 
+  
 
 ## Simulation
 

pyrddlgym_jax-2.2/pyRDDLGym_jax/__init__.py

@@ -0,0 +1 @@
+__version__ = '2.2'

{pyrddlgym_jax-2.1 → pyrddlgym_jax-2.2}/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
@@ -1212,17 +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
@@ -1231,6 +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__()
 
@@ -1238,8 +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:
@@ -1249,36 +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)
@@ -1289,7 +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)
 
@@ -1299,19 +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)
@@ -1321,14 +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)
@@ -1336,37 +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]
@@ -1384,42 +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)
+                    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
@@ -1462,14 +1546,14 @@ def mean_deviation_utility(returns: jnp.ndarray, beta: float) -> float:
 @jax.jit
 def mean_semideviation_utility(returns: jnp.ndarray, beta: float) -> float:
     mu = jnp.mean(returns)
-    msd = jnp.sqrt(jnp.mean(jnp.minimum(0.0, returns - mu) ** 2))
+    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.minimum(0.0, returns - mu) ** 2)
+    msv = jnp.mean(jnp.square(jnp.minimum(0.0, returns - mu)))
     return mu - 0.5 * beta * msv
 
 
@@ -1768,7 +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:
@@ -1831,6 +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,
@@ -1855,23 +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
@@ -2175,11 +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:
@@ -2191,18 +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, 
-                         bar_format='{l_bar}{bar}| {elapsed} {postfix}', 
-                         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
@@ -2213,8 +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)
@@ -2224,8 +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)
@@ -2252,7 +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
@@ -2311,14 +2393,15 @@ r"""
             
             # 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',
                     refresh=False
                 )
-                iters.set_postfix_str(f"{(it + 1) / elapsed:.2f}it/s", refresh=True)
+                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:
@@ -2339,7 +2422,7 @@ r"""
 
         # release resources
         if print_progress:
-            iters.close()
+            progress_bar.close()
         
         # validate the test return
         if log:

{pyrddlgym_jax-2.1 → pyrddlgym_jax-2.2}/pyRDDLGym_jax.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: pyRDDLGym-jax
-Version: 2.1
+Version: 2.2
 Summary: pyRDDLGym-jax: automatic differentiation for solving sequential planning problems in JAX.
 Home-page: https://github.com/pyrddlgym-project/pyRDDLGym-jax
 Author: Michael Gimelfarb, Ayal Taitler, Scott Sanner
@@ -58,8 +58,11 @@ Dynamic: summary
 
 Purpose:
 
-1. automatic translation of any RDDL description file into a differentiable simulator in JAX
-2. flexible policy class representations, automatic model relaxations for working in discrete and hybrid domains, and Bayesian hyper-parameter tuning.
+1. automatic translation of RDDL description files into differentiable JAX simulators
+2. implementation of (highly configurable) operator relaxations for working in discrete and hybrid domains
+3. flexible policy representations and automated Bayesian hyper-parameter tuning
+4. interactive dashboard for dyanmic visualization and debugging
+5. hybridization with parameter-exploring policy gradients.
 
 Some demos of solved problems by JaxPlan:
 
@@ -235,8 +238,23 @@ More documentation about this and other new features will be coming soon.
 
 ## Tuning the Planner
 
-It is easy to tune the planner's hyper-parameters efficiently and automatically using Bayesian optimization.
-To do this, first create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
+A basic run script is provided to run automatic Bayesian hyper-parameter tuning for the most sensitive parameters of JaxPlan:
+
+```shell
+jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
+```
+
+where:
+- ``domain`` is the domain identifier as specified in rddlrepository
+- ``instance`` is the instance identifier
+- ``method`` is the planning method to use (i.e. drp, slp, replan)
+- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
+- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
+- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
+- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
+
+It is easy to tune a custom range of the planner's hyper-parameters efficiently. 
+First create a config file template with patterns replacing concrete parameter values that you want to tune, e.g.:
 
 ```ini
 [Model]
@@ -260,7 +278,7 @@ train_on_reset=True
 
 would allow to tune the sharpness of model relaxations, and the learning rate of the optimizer.
 
-Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand:
+Next, you must link the patterns in the config with concrete hyper-parameter ranges the tuner will understand, and run the optimizer:
 
 ```python
 import pyRDDLGym
@@ -292,22 +310,7 @@ tuning = JaxParameterTuning(env=env,
                             gp_iters=iters)
 tuning.tune(key=42, log_file='path/to/log.csv')
 ```
- 
-A basic run script is provided to run the automatic hyper-parameter tuning for the most sensitive parameters of JaxPlan:
-
-```shell
-jaxplan tune <domain> <instance> <method> <trials> <iters> <workers> <dashboard>
-```
-
-where:
-- ``domain`` is the domain identifier as specified in rddlrepository
-- ``instance`` is the instance identifier
-- ``method`` is the planning method to use (i.e. drp, slp, replan)
-- ``trials`` is the (optional) number of trials/episodes to average in evaluating each hyper-parameter setting
-- ``iters`` is the (optional) maximum number of iterations/evaluations of Bayesian optimization to perform
-- ``workers`` is the (optional) number of parallel evaluations to be done at each iteration, e.g. the total evaluations = ``iters * workers``
-- ``dashboard`` is whether the optimizations are tracked in the dashboard application.
- 
+  
 
 ## Simulation
 

{pyrddlgym_jax-2.1 → pyrddlgym_jax-2.2}/setup.py

@@ -19,7 +19,7 @@ long_description = (Path(__file__).parent / "README.md").read_text()
 
 setup(
       name='pyRDDLGym-jax',
-      version='2.1',
+      version='2.2',
       author="Michael Gimelfarb, Ayal Taitler, Scott Sanner",
       author_email="mike.gimelfarb@mail.utoronto.ca, ataitler@gmail.com, ssanner@mie.utoronto.ca",
       description="pyRDDLGym-jax: automatic differentiation for solving sequential planning problems in JAX.",

pyrddlgym_jax-2.1/pyRDDLGym_jax/__init__.py

@@ -1 +0,0 @@
-__version__ = '2.1'