pyRDDLGym-jax 0.1__py3-none-any.whl → 0.2__py3-none-any.whl

pyRDDLGym_jax/core/compiler.py

@@ -4,7 +4,7 @@ import jax.numpy as jnp
 import jax.random as random
 import jax.scipy as scipy 
 import traceback
-from typing import Callable, Dict, List
+from typing import Any, Callable, Dict, List, Optional
 
 from pyRDDLGym.core.debug.exception import raise_warning
 
@@ -13,8 +13,9 @@ try:
     from tensorflow_probability.substrates import jax as tfp
 except Exception:
     raise_warning('Failed to import tensorflow-probability: '
-                  'compilation of some complex distributions will not work.', 
-                  'red')
+                  'compilation of some complex distributions '
+                  '(Binomial, Negative-Binomial, Multinomial) will fail. '
+                  'Please ensure this package is installed correctly.', 'red')
     traceback.print_exc()
     tfp = None
 
@@ -32,6 +33,98 @@ from pyRDDLGym.core.debug.logger import Logger
 from pyRDDLGym.core.simulator import RDDLSimulatorPrecompiled
 
 
+# ===========================================================================
+# EXACT RDDL TO JAX COMPILATION RULES
+# ===========================================================================
+     
+def _function_unary_exact_named(op, name):
+        
+    def _jax_wrapped_unary_fn_exact(x, param):
+        return op(x)
+        
+    return _jax_wrapped_unary_fn_exact
+        
+        
+def _function_unary_exact_named_gamma():
+        
+    def _jax_wrapped_unary_gamma_exact(x, param):
+        return jnp.exp(scipy.special.gammaln(x))
+        
+    return _jax_wrapped_unary_gamma_exact        
+    
+    
+def _function_binary_exact_named(op, name):
+        
+    def _jax_wrapped_binary_fn_exact(x, y, param):
+        return op(x, y)
+        
+    return _jax_wrapped_binary_fn_exact
+    
+    
+def _function_binary_exact_named_implies():
+        
+    def _jax_wrapped_binary_implies_exact(x, y, param):
+        return jnp.logical_or(jnp.logical_not(x), y)
+        
+    return _jax_wrapped_binary_implies_exact
+    
+    
+def _function_binary_exact_named_log():
+        
+    def _jax_wrapped_binary_log_exact(x, y, param):
+        return jnp.log(x) / jnp.log(y)
+        
+    return _jax_wrapped_binary_log_exact
+    
+    
+def _function_aggregation_exact_named(op, name):
+        
+    def _jax_wrapped_aggregation_fn_exact(x, axis, param):
+        return op(x, axis=axis)
+        
+    return _jax_wrapped_aggregation_fn_exact
+    
+    
+def _function_if_exact_named():
+        
+    def _jax_wrapped_if_exact(c, a, b, param):
+        return jnp.where(c, a, b)
+        
+    return _jax_wrapped_if_exact
+    
+    
+def _function_switch_exact_named():
+        
+    def _jax_wrapped_switch_exact(pred, cases, param):
+        pred = pred[jnp.newaxis, ...]
+        sample = jnp.take_along_axis(cases, pred, axis=0)
+        assert sample.shape[0] == 1
+        return sample[0, ...]
+
+    return _jax_wrapped_switch_exact
+    
+    
+def _function_bernoulli_exact_named():
+        
+    def _jax_wrapped_bernoulli_exact(key, prob, param):
+        return random.bernoulli(key, prob)
+        
+    return _jax_wrapped_bernoulli_exact
+    
+    
+def _function_discrete_exact_named():
+        
+    def _jax_wrapped_discrete_exact(key, prob, param):
+        logits = jnp.log(prob)
+        sample = random.categorical(key=key, logits=logits, axis=-1)
+        out_of_bounds = jnp.logical_not(jnp.logical_and(
+            jnp.all(prob >= 0),
+            jnp.allclose(jnp.sum(prob, axis=-1), 1.0)))
+        return sample, out_of_bounds
+        
+    return _jax_wrapped_discrete_exact
+
+
 class JaxRDDLCompiler:
     '''Compiles a RDDL AST representation into an equivalent JAX representation.
     All operations are identical to their numpy equivalents.
@@ -39,10 +132,97 @@ class JaxRDDLCompiler:
     
     MODEL_PARAM_TAG_SEPARATOR = '___'
     
+    # ===========================================================================
+    # EXACT RDDL TO JAX COMPILATION RULES BY DEFAULT
+    # ===========================================================================
+
+    EXACT_RDDL_TO_JAX_NEGATIVE = _function_unary_exact_named(jnp.negative, 'negative')
+    
+    EXACT_RDDL_TO_JAX_ARITHMETIC = {
+        '+': _function_binary_exact_named(jnp.add, 'add'),
+        '-': _function_binary_exact_named(jnp.subtract, 'subtract'),
+        '*': _function_binary_exact_named(jnp.multiply, 'multiply'),
+        '/': _function_binary_exact_named(jnp.divide, 'divide')
+    }    
+        
+    EXACT_RDDL_TO_JAX_RELATIONAL = {
+        '>=': _function_binary_exact_named(jnp.greater_equal, 'greater_equal'),
+        '<=': _function_binary_exact_named(jnp.less_equal, 'less_equal'),
+        '<': _function_binary_exact_named(jnp.less, 'less'),
+        '>': _function_binary_exact_named(jnp.greater, 'greater'),
+        '==': _function_binary_exact_named(jnp.equal, 'equal'),
+        '~=': _function_binary_exact_named(jnp.not_equal, 'not_equal')
+    }
+        
+    EXACT_RDDL_TO_JAX_LOGICAL = {
+        '^': _function_binary_exact_named(jnp.logical_and, 'and'),
+        '&': _function_binary_exact_named(jnp.logical_and, 'and'),
+        '|': _function_binary_exact_named(jnp.logical_or, 'or'),
+        '~': _function_binary_exact_named(jnp.logical_xor, 'xor'),
+        '=>': _function_binary_exact_named_implies(),
+        '<=>': _function_binary_exact_named(jnp.equal, 'iff')
+    }
+    
+    EXACT_RDDL_TO_JAX_LOGICAL_NOT = _function_unary_exact_named(jnp.logical_not, 'not')
+    
+    EXACT_RDDL_TO_JAX_AGGREGATION = {
+        'sum': _function_aggregation_exact_named(jnp.sum, 'sum'),
+        'avg': _function_aggregation_exact_named(jnp.mean, 'avg'),
+        'prod': _function_aggregation_exact_named(jnp.prod, 'prod'),
+        'minimum': _function_aggregation_exact_named(jnp.min, 'minimum'),
+        'maximum': _function_aggregation_exact_named(jnp.max, 'maximum'),
+        'forall': _function_aggregation_exact_named(jnp.all, 'forall'),
+        'exists': _function_aggregation_exact_named(jnp.any, 'exists'),
+        'argmin': _function_aggregation_exact_named(jnp.argmin, 'argmin'),
+        'argmax': _function_aggregation_exact_named(jnp.argmax, 'argmax')
+    }
+    
+    EXACT_RDDL_TO_JAX_UNARY = {        
+        'abs': _function_unary_exact_named(jnp.abs, 'abs'),
+        'sgn': _function_unary_exact_named(jnp.sign, 'sgn'),
+        'round': _function_unary_exact_named(jnp.round, 'round'),
+        'floor': _function_unary_exact_named(jnp.floor, 'floor'),
+        'ceil': _function_unary_exact_named(jnp.ceil, 'ceil'),
+        'cos': _function_unary_exact_named(jnp.cos, 'cos'),
+        'sin': _function_unary_exact_named(jnp.sin, 'sin'),
+        'tan': _function_unary_exact_named(jnp.tan, 'tan'),
+        'acos': _function_unary_exact_named(jnp.arccos, 'acos'),
+        'asin': _function_unary_exact_named(jnp.arcsin, 'asin'),
+        'atan': _function_unary_exact_named(jnp.arctan, 'atan'),
+        'cosh': _function_unary_exact_named(jnp.cosh, 'cosh'),
+        'sinh': _function_unary_exact_named(jnp.sinh, 'sinh'),
+        'tanh': _function_unary_exact_named(jnp.tanh, 'tanh'),
+        'exp': _function_unary_exact_named(jnp.exp, 'exp'),
+        'ln': _function_unary_exact_named(jnp.log, 'ln'),
+        'sqrt': _function_unary_exact_named(jnp.sqrt, 'sqrt'),
+        'lngamma': _function_unary_exact_named(scipy.special.gammaln, 'lngamma'),
+        'gamma': _function_unary_exact_named_gamma()
+    }      
+        
+    EXACT_RDDL_TO_JAX_BINARY = {
+        'div': _function_binary_exact_named(jnp.floor_divide, 'div'),
+        'mod': _function_binary_exact_named(jnp.mod, 'mod'),
+        'fmod': _function_binary_exact_named(jnp.mod, 'fmod'),
+        'min': _function_binary_exact_named(jnp.minimum, 'min'),
+        'max': _function_binary_exact_named(jnp.maximum, 'max'),
+        'pow': _function_binary_exact_named(jnp.power, 'pow'),
+        'log': _function_binary_exact_named_log(),
+        'hypot': _function_binary_exact_named(jnp.hypot, 'hypot'),
+    }
+    
+    EXACT_RDDL_TO_JAX_IF = _function_if_exact_named()
+    
+    EXACT_RDDL_TO_JAX_SWITCH = _function_switch_exact_named()
+    
+    EXACT_RDDL_TO_JAX_BERNOULLI = _function_bernoulli_exact_named()
+    
+    EXACT_RDDL_TO_JAX_DISCRETE = _function_discrete_exact_named()
+
     def __init__(self, rddl: RDDLLiftedModel,
                  allow_synchronous_state: bool=True,
-                 logger: Logger=None,
-                 use64bit: bool=False) -> None:
+                 logger: Optional[Logger]=None,
+                 use64bit: bool=False,
+                 compile_non_fluent_exact: bool=True) -> None:
         '''Creates a new RDDL to Jax compiler.
         
         :param rddl: the RDDL model to compile into Jax
@@ -50,11 +230,14 @@ class JaxRDDLCompiler:
         on each other
         :param logger: to log information about compilation to file
         :param use64bit: whether to use 64 bit arithmetic
+        :param compile_non_fluent_exact: whether non-fluent expressions 
+        are always compiled using exact JAX expressions.
         '''
         self.rddl = rddl
         self.logger = logger
         # jax.config.update('jax_log_compiles', True) # for testing ONLY
         
+        self.use64bit = use64bit
         if use64bit:
             self.INT = jnp.int64
             self.REAL = jnp.float64
@@ -62,6 +245,7 @@ class JaxRDDLCompiler:
         else:
             self.INT = jnp.int32
             self.REAL = jnp.float32
+            jax.config.update('jax_enable_x64', False)
         self.ONE = jnp.asarray(1, dtype=self.INT)
         self.JAX_TYPES = {
             'int': self.INT,
@@ -70,17 +254,16 @@ class JaxRDDLCompiler:
         }
         
         # compile initial values
-        if self.logger is not None:
-            self.logger.clear()
-        initializer = RDDLValueInitializer(rddl, logger=self.logger)
+        initializer = RDDLValueInitializer(rddl)
         self.init_values = initializer.initialize()
         
         # compute dependency graph for CPFs and sort them by evaluation order
-        sorter = RDDLLevelAnalysis(rddl, allow_synchronous_state, logger=self.logger)
+        sorter = RDDLLevelAnalysis(
+            rddl, allow_synchronous_state=allow_synchronous_state)
         self.levels = sorter.compute_levels()        
         
         # trace expressions to cache information to be used later
-        tracer = RDDLObjectsTracer(rddl, logger=self.logger, cpf_levels=self.levels)
+        tracer = RDDLObjectsTracer(rddl, cpf_levels=self.levels)
         self.traced = tracer.trace()
         
         # extract the box constraints on actions
@@ -92,92 +275,42 @@ class JaxRDDLCompiler:
         constraints = RDDLConstraints(simulator, vectorized=True)
         self.constraints = constraints
         
-        # basic operations
-        self.NEGATIVE = lambda x, param: jnp.negative(x)  
-        self.ARITHMETIC_OPS = {
-            '+': lambda x, y, param: jnp.add(x, y),
-            '-': lambda x, y, param: jnp.subtract(x, y),
-            '*': lambda x, y, param: jnp.multiply(x, y),
-            '/': lambda x, y, param: jnp.divide(x, y)
-        }    
-        self.RELATIONAL_OPS = {
-            '>=': lambda x, y, param: jnp.greater_equal(x, y),
-            '<=': lambda x, y, param: jnp.less_equal(x, y),
-            '<': lambda x, y, param: jnp.less(x, y),
-            '>': lambda x, y, param: jnp.greater(x, y),
-            '==': lambda x, y, param: jnp.equal(x, y),
-            '~=': lambda x, y, param: jnp.not_equal(x, y)
-        }
-        self.LOGICAL_NOT = lambda x, param: jnp.logical_not(x)
-        self.LOGICAL_OPS = {
-            '^': lambda x, y, param: jnp.logical_and(x, y),
-            '&': lambda x, y, param: jnp.logical_and(x, y),
-            '|': lambda x, y, param: jnp.logical_or(x, y),
-            '~': lambda x, y, param: jnp.logical_xor(x, y),
-            '=>': lambda x, y, param: jnp.logical_or(jnp.logical_not(x), y),
-            '<=>': lambda x, y, param: jnp.equal(x, y)
-        }
-        self.AGGREGATION_OPS = {
-            'sum': lambda x, axis, param: jnp.sum(x, axis=axis),
-            'avg': lambda x, axis, param: jnp.mean(x, axis=axis),
-            'prod': lambda x, axis, param: jnp.prod(x, axis=axis),
-            'minimum': lambda x, axis, param: jnp.min(x, axis=axis),
-            'maximum': lambda x, axis, param: jnp.max(x, axis=axis),
-            'forall': lambda x, axis, param: jnp.all(x, axis=axis),
-            'exists': lambda x, axis, param: jnp.any(x, axis=axis),
-            'argmin': lambda x, axis, param: jnp.argmin(x, axis=axis),
-            'argmax': lambda x, axis, param: jnp.argmax(x, axis=axis)
-        }
+        # basic operations - these can be override in subclasses
+        self.compile_non_fluent_exact = compile_non_fluent_exact
+        self.NEGATIVE = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_NEGATIVE
+        self.ARITHMETIC_OPS = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_ARITHMETIC.copy()
+        self.RELATIONAL_OPS = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_RELATIONAL.copy()
+        self.LOGICAL_NOT = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_LOGICAL_NOT
+        self.LOGICAL_OPS = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_LOGICAL.copy()
+        self.AGGREGATION_OPS = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_AGGREGATION.copy()
         self.AGGREGATION_BOOL = {'forall', 'exists'}
-        self.KNOWN_UNARY = {        
-            'abs': lambda x, param: jnp.abs(x),
-            'sgn': lambda x, param: jnp.sign(x),
-            'round': lambda x, param: jnp.round(x),
-            'floor': lambda x, param: jnp.floor(x),
-            'ceil': lambda x, param: jnp.ceil(x),
-            'cos': lambda x, param: jnp.cos(x),
-            'sin': lambda x, param: jnp.sin(x),
-            'tan': lambda x, param: jnp.tan(x),
-            'acos': lambda x, param: jnp.arccos(x),
-            'asin': lambda x, param: jnp.arcsin(x),
-            'atan': lambda x, param: jnp.arctan(x),
-            'cosh': lambda x, param: jnp.cosh(x),
-            'sinh': lambda x, param: jnp.sinh(x),
-            'tanh': lambda x, param: jnp.tanh(x),
-            'exp': lambda x, param: jnp.exp(x),
-            'ln': lambda x, param: jnp.log(x),
-            'sqrt': lambda x, param: jnp.sqrt(x),
-            'lngamma': lambda x, param: scipy.special.gammaln(x),
-            'gamma': lambda x, param: jnp.exp(scipy.special.gammaln(x))
-        }        
-        self.KNOWN_BINARY = {
-            'div': lambda x, y, param: jnp.floor_divide(x, y),
-            'mod': lambda x, y, param: jnp.mod(x, y),
-            'fmod': lambda x, y, param: jnp.mod(x, y),
-            'min': lambda x, y, param: jnp.minimum(x, y),
-            'max': lambda x, y, param: jnp.maximum(x, y),
-            'pow': lambda x, y, param: jnp.power(x, y),
-            'log': lambda x, y, param: jnp.log(x) / jnp.log(y),
-            'hypot': lambda x, y, param: jnp.hypot(x, y)
-        }
-        
+        self.KNOWN_UNARY = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_UNARY.copy()
+        self.KNOWN_BINARY = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_BINARY.copy()
+        self.IF_HELPER = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_IF
+        self.SWITCH_HELPER = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_SWITCH
+        self.BERNOULLI_HELPER = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_BERNOULLI
+        self.DISCRETE_HELPER = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_DISCRETE
+    
     # ===========================================================================
     # main compilation subroutines
     # ===========================================================================
      
-    def compile(self, log_jax_expr: bool=False) -> None: 
+    def compile(self, log_jax_expr: bool=False, heading: str='') -> None: 
         '''Compiles the current RDDL into Jax expressions.
         
         :param log_jax_expr: whether to pretty-print the compiled Jax functions
         to the log file
+        :param heading: the heading to print before compilation information
         '''
-        info = {}
+        info = ({}, [])
         self.invariants = self._compile_constraints(self.rddl.invariants, info)
         self.preconditions = self._compile_constraints(self.rddl.preconditions, info)
         self.terminations = self._compile_constraints(self.rddl.terminations, info)
         self.cpfs = self._compile_cpfs(info)
         self.reward = self._compile_reward(info)
-        self.model_params = info
+        self.model_params = {key: value 
+                             for (key, (value, *_)) in info[0].items()}
+        self.relaxations = info[1]
         
         if log_jax_expr and self.logger is not None:
             printed = self.print_jax()
@@ -189,6 +322,7 @@ class JaxRDDLCompiler:
             printed_terminals = '\n\n'.join(v for v in printed['terminations'])
             printed_params = '\n'.join(f'{k}: {v}' for (k, v) in info.items())
             message = (
+                f'[info] {heading}\n'
                 f'[info] compiled JAX CPFs:\n\n'
                 f'{printed_cpfs}\n\n'
                 f'[info] compiled JAX reward:\n\n'
@@ -281,17 +415,21 @@ class JaxRDDLCompiler:
         return jax_inequalities, jax_equalities
     
     def compile_transition(self, check_constraints: bool=False,
-                           constraint_func: bool=False):
+                           constraint_func: bool=False) -> Callable:
         '''Compiles the current RDDL model into a JAX transition function that 
         samples the next state.
         
-        The signature of the returned function is (key, actions, subs, 
-        model_params), where:
+        The arguments of the returned function is:
             - key is the PRNG key
             - actions is the dict of action tensors
             - subs is the dict of current pvar value tensors
             - model_params is a dict of parameters for the relaxed model.
-                    
+        
+        The returned value of the function is:
+            - subs is the returned next epoch fluent values
+            - log includes all the auxiliary information about constraints 
+              satisfied, errors, etc.
+            
         constraint_func provides the option to compile nonlinear constraints:
         
             1. f(s, a) ?? g(s, a)
@@ -361,6 +499,10 @@ class JaxRDDLCompiler:
             reward, key, err = reward_fn(subs, model_params, key)
             errors |= err
             
+            # calculate fluent values
+            fluents = {name: values for (name, values) in subs.items() 
+                       if name not in rddl.non_fluents}
+            
             # set the next state to the current state
             for (state, next_state) in rddl.next_state.items():
                 subs[state] = subs[next_state]
@@ -383,8 +525,7 @@ class JaxRDDLCompiler:
             
             # prepare the return value
             log = {
-                'pvar': subs,
-                'action': actions,
+                'fluents': fluents,
                 'reward': reward,
                 'error': errors,
                 'precondition': precond_check,
@@ -395,7 +536,7 @@ class JaxRDDLCompiler:
                 log['inequalities'] = inequalities
                 log['equalities'] = equalities
                 
-            return log
+            return subs, log
         
         return _jax_wrapped_single_step        
     
@@ -403,18 +544,28 @@ class JaxRDDLCompiler:
                          n_steps: int,
                          n_batch: int,
                          check_constraints: bool=False,
-                         constraint_func: bool=False):
+                         constraint_func: bool=False) -> Callable:
         '''Compiles the current RDDL model into a JAX transition function that 
         samples trajectories with a fixed horizon from a policy.
         
-        The signature of the policy function is (key, params, hyperparams, 
-        step, states), where:
+        The arguments of the returned function is:
+            - key is the PRNG key (used by a stochastic policy)
+            - policy_params is a pytree of trainable policy weights
+            - hyperparams is a pytree of (optional) fixed policy hyper-parameters
+            - subs is the dictionary of current fluent tensor values
+            - model_params is a dict of model hyperparameters.
+        
+        The returned value of the returned function is:
+            - log is the dictionary of all trajectory information, including
+              constraints that were satisfied, errors, etc. 
+            
+        The arguments of the policy function is:
             - key is the PRNG key (used by a stochastic policy)
             - params is a pytree of trainable policy weights
             - hyperparams is a pytree of (optional) fixed policy hyper-parameters
             - step is the time index of the decision in the current rollout
             - states is a dict of tensors for the current observation.
-            
+        
         :param policy: a Jax compiled function for the policy as described above
         decision epoch, state dict, and an RNG key and returns an action dict
         :param n_steps: the rollout horizon
@@ -428,27 +579,32 @@ class JaxRDDLCompiler:
         rddl = self.rddl
         jax_step_fn = self.compile_transition(check_constraints, constraint_func)
         
+        # for POMDP only observ-fluents are assumed visible to the policy
+        if rddl.observ_fluents:
+            observed_vars = rddl.observ_fluents
+        else:
+            observed_vars = rddl.state_fluents
+            
         # evaluate the step from the policy
         def _jax_wrapped_single_step_policy(key, policy_params, hyperparams, 
                                             step, subs, model_params):
             states = {var: values 
                       for (var, values) in subs.items()
-                      if rddl.variable_types[var] == 'state-fluent'}
+                      if var in observed_vars}
             actions = policy(key, policy_params, hyperparams, step, states)
             key, subkey = random.split(key)
-            log = jax_step_fn(subkey, actions, subs, model_params)
-            return log
+            subs, log = jax_step_fn(subkey, actions, subs, model_params)
+            return subs, log
                         
         # do a batched step update from the policy
         def _jax_wrapped_batched_step_policy(carry, step):
             key, policy_params, hyperparams, subs, model_params = carry  
             key, *subkeys = random.split(key, num=1 + n_batch)
             keys = jnp.asarray(subkeys)
-            log = jax.vmap(
+            subs, log = jax.vmap(
                 _jax_wrapped_single_step_policy,
                 in_axes=(0, None, None, None, 0, None)
             )(keys, policy_params, hyperparams, step, subs, model_params)
-            subs = log['pvar']
             carry = (key, policy_params, hyperparams, subs, model_params)
             return carry, log            
             
@@ -467,7 +623,7 @@ class JaxRDDLCompiler:
     # error checks
     # ===========================================================================
     
-    def print_jax(self) -> Dict[str, object]:
+    def print_jax(self) -> Dict[str, Any]:
         '''Returns a dictionary containing the string representations of all 
         Jax compiled expressions from the RDDL file.
         '''
@@ -564,7 +720,7 @@ class JaxRDDLCompiler:
     }
     
     @staticmethod
-    def get_error_codes(error):
+    def get_error_codes(error: int) -> List[int]:
         '''Given a compacted integer error flag from the execution of Jax, and 
         decomposes it into individual error codes.
         '''
@@ -573,7 +729,7 @@ class JaxRDDLCompiler:
         return errors
     
     @staticmethod
-    def get_error_messages(error):
+    def get_error_messages(error: int) -> List[str]:
         '''Given a compacted integer error flag from the execution of Jax, and 
         decomposes it into error strings.
         '''
@@ -586,28 +742,40 @@ class JaxRDDLCompiler:
     # ===========================================================================
     
     def _unwrap(self, op, expr_id, info):
-        sep = JaxRDDLCompiler.MODEL_PARAM_TAG_SEPARATOR
         jax_op, name = op, None
+        model_params, relaxed_list = info
         if isinstance(op, tuple):
             jax_op, param = op
             if param is not None:
                 tags, values = param
+                sep = JaxRDDLCompiler.MODEL_PARAM_TAG_SEPARATOR
                 if isinstance(tags, tuple):
                     name = sep.join(tags)
                 else:
                     name = str(tags)
                 name = f'{name}{sep}{expr_id}'
-                if name in info:
-                    raise Exception(f'Model parameter {name} is already defined.')
-                info[name] = values
+                if name in model_params:
+                    raise RuntimeError(
+                        f'Internal error: model parameter {name} is already defined.')
+                model_params[name] = (values, tags, expr_id, jax_op.__name__)
+            relaxed_list.append((param, expr_id, jax_op.__name__))
         return jax_op, name
     
-    def get_ids_of_parameterized_expressions(self) -> List[int]:
-        '''Returns a list of expression IDs that have tuning parameters.'''
-        sep = JaxRDDLCompiler.MODEL_PARAM_TAG_SEPARATOR
-        ids = [int(key.split(sep)[-1]) for key in self.model_params] 
-        return ids
-    
+    def summarize_model_relaxations(self) -> str:
+        '''Returns a string of information about model relaxations in the
+        compiled model.'''
+        occurence_by_type = {}
+        for (_, expr_id, jax_op) in self.relaxations:
+            etype = self.traced.lookup(expr_id).etype
+            source = f'{etype[1]} ({etype[0]})'
+            sub = f'{source:<30} --> {jax_op}'
+            occurence_by_type[sub] = occurence_by_type.get(sub, 0) + 1        
+        col = "{:<80} {:<10}\n"
+        table = col.format('Substitution', 'Count')
+        for (sub, occurs) in occurence_by_type.items():
+            table += col.format(sub, occurs)
+        return table
+        
     # ===========================================================================
     # expression compilation
     # ===========================================================================
@@ -640,7 +808,8 @@ class JaxRDDLCompiler:
             raise RDDLNotImplementedError(
                 f'Internal error: expression type {expr} is not supported.\n' + 
                 print_stack_trace(expr))
-                
+            
+        # force type cast of tensor as required by caller
         if dtype is not None:
             jax_expr = self._jax_cast(jax_expr, dtype)
         
@@ -660,6 +829,17 @@ class JaxRDDLCompiler:
         
         return _jax_wrapped_cast
    
+    def _fix_dtype(self, value):
+        dtype = jnp.atleast_1d(value).dtype
+        if jnp.issubdtype(dtype, jnp.integer):
+            return self.INT
+        elif jnp.issubdtype(dtype, jnp.floating):
+            return self.REAL
+        elif jnp.issubdtype(dtype, jnp.bool_) or jnp.issubdtype(dtype, bool):
+            return bool
+        else:
+            raise TypeError(f'Invalid type {dtype} of {value}.')
+       
     # ===========================================================================
     # leaves
     # ===========================================================================
@@ -669,7 +849,7 @@ class JaxRDDLCompiler:
         cached_value = self.traced.cached_sim_info(expr)
         
         def _jax_wrapped_constant(x, params, key):
-            sample = jnp.asarray(cached_value)
+            sample = jnp.asarray(cached_value, dtype=self._fix_dtype(cached_value))
             return sample, key, NORMAL
 
         return _jax_wrapped_constant
@@ -693,7 +873,7 @@ class JaxRDDLCompiler:
             cached_value = cached_info
 
             def _jax_wrapped_object(x, params, key):
-                sample = jnp.asarray(cached_value)
+                sample = jnp.asarray(cached_value, dtype=self._fix_dtype(cached_value))
                 return sample, key, NORMAL
             
             return _jax_wrapped_object
@@ -702,7 +882,8 @@ class JaxRDDLCompiler:
         elif cached_info is None:
             
             def _jax_wrapped_pvar_scalar(x, params, key):
-                sample = jnp.asarray(x[var])
+                value = x[var]
+                sample = jnp.asarray(value, dtype=self._fix_dtype(value))
                 return sample, key, NORMAL
             
             return _jax_wrapped_pvar_scalar
@@ -721,7 +902,8 @@ class JaxRDDLCompiler:
                 
                 def _jax_wrapped_pvar_tensor_nested(x, params, key):
                     error = NORMAL
-                    sample = jnp.asarray(x[var])
+                    value = x[var]
+                    sample = jnp.asarray(value, dtype=self._fix_dtype(value))
                     new_slices = [None] * len(jax_nested_expr)
                     for (i, jax_expr) in enumerate(jax_nested_expr):
                         new_slices[i], key, err = jax_expr(x, params, key)
@@ -736,7 +918,8 @@ class JaxRDDLCompiler:
             else:
     
                 def _jax_wrapped_pvar_tensor_non_nested(x, params, key):
-                    sample = jnp.asarray(x[var])
+                    value = x[var]
+                    sample = jnp.asarray(value, dtype=self._fix_dtype(value))
                     if slices:
                         sample = sample[slices]
                     if axis:
@@ -795,16 +978,23 @@ class JaxRDDLCompiler:
     
     def _jax_arithmetic(self, expr, info):
         _, op = expr.etype
-        valid_ops = self.ARITHMETIC_OPS
+        
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):
+            valid_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_ARITHMETIC
+            negative_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_NEGATIVE
+        else:
+            valid_ops = self.ARITHMETIC_OPS
+            negative_op = self.NEGATIVE            
         JaxRDDLCompiler._check_valid_op(expr, valid_ops)
-                    
+        
+        # recursively compile arguments
         args = expr.args
         n = len(args)
-        
         if n == 1 and op == '-':
             arg, = args
             jax_expr = self._jax(arg, info)
-            jax_op, jax_param = self._unwrap(self.NEGATIVE, expr.id, info)
+            jax_op, jax_param = self._unwrap(negative_op, expr.id, info)
             return self._jax_unary(jax_expr, jax_op, jax_param, at_least_int=True)
                     
         elif n == 2:
@@ -819,29 +1009,42 @@ class JaxRDDLCompiler:
     
     def _jax_relational(self, expr, info):
         _, op = expr.etype
-        valid_ops = self.RELATIONAL_OPS
+        
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):
+            valid_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_RELATIONAL
+        else:
+            valid_ops = self.RELATIONAL_OPS
         JaxRDDLCompiler._check_valid_op(expr, valid_ops)
-        JaxRDDLCompiler._check_num_args(expr, 2)
+        jax_op, jax_param = self._unwrap(valid_ops[op], expr.id, info)
         
+        # recursively compile arguments
+        JaxRDDLCompiler._check_num_args(expr, 2)
         lhs, rhs = expr.args
         jax_lhs = self._jax(lhs, info)
         jax_rhs = self._jax(rhs, info)
-        jax_op, jax_param = self._unwrap(valid_ops[op], expr.id, info)
         return self._jax_binary(
             jax_lhs, jax_rhs, jax_op, jax_param, at_least_int=True)
            
     def _jax_logical(self, expr, info):
         _, op = expr.etype
-        valid_ops = self.LOGICAL_OPS    
+        
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):
+            valid_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_LOGICAL  
+            logical_not_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_LOGICAL_NOT
+        else:
+            valid_ops = self.LOGICAL_OPS 
+            logical_not_op = self.LOGICAL_NOT 
         JaxRDDLCompiler._check_valid_op(expr, valid_ops)
                 
+        # recursively compile arguments
         args = expr.args
-        n = len(args)
-        
+        n = len(args)        
         if n == 1 and op == '~':
             arg, = args
             jax_expr = self._jax(arg, info)
-            jax_op, jax_param = self._unwrap(self.LOGICAL_NOT, expr.id, info)
+            jax_op, jax_param = self._unwrap(logical_not_op, expr.id, info)
             return self._jax_unary(jax_expr, jax_op, jax_param, check_dtype=bool)
         
         elif n == 2:
@@ -856,17 +1059,21 @@ class JaxRDDLCompiler:
     
     def _jax_aggregation(self, expr, info):
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_CAST']
-        
         _, op = expr.etype
-        valid_ops = self.AGGREGATION_OPS      
+        
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):
+            valid_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_AGGREGATION   
+        else:
+            valid_ops = self.AGGREGATION_OPS   
         JaxRDDLCompiler._check_valid_op(expr, valid_ops) 
-        is_floating = op not in self.AGGREGATION_BOOL
+        jax_op, jax_param = self._unwrap(valid_ops[op], expr.id, info)
         
+        # recursively compile arguments
+        is_floating = op not in self.AGGREGATION_BOOL
         * _, arg = expr.args  
-        _, axes = self.traced.cached_sim_info(expr)
-        
+        _, axes = self.traced.cached_sim_info(expr)        
         jax_expr = self._jax(arg, info)
-        jax_op, jax_param = self._unwrap(valid_ops[op], expr.id, info)
         
         def _jax_wrapped_aggregation(x, params, key):
             sample, key, err = jax_expr(x, params, key)
@@ -884,21 +1091,28 @@ class JaxRDDLCompiler:
     def _jax_functional(self, expr, info):
         _, op = expr.etype
         
-        # unary function
-        if op in self.KNOWN_UNARY:
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):            
+            unary_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_UNARY
+            binary_ops = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_BINARY
+        else:
+            unary_ops = self.KNOWN_UNARY
+            binary_ops = self.KNOWN_BINARY
+        
+        # recursively compile arguments
+        if op in unary_ops:
             JaxRDDLCompiler._check_num_args(expr, 1)                            
             arg, = expr.args
             jax_expr = self._jax(arg, info)
-            jax_op, jax_param = self._unwrap(self.KNOWN_UNARY[op], expr.id, info)
+            jax_op, jax_param = self._unwrap(unary_ops[op], expr.id, info)
             return self._jax_unary(jax_expr, jax_op, jax_param, at_least_int=True)
             
-        # binary function
-        elif op in self.KNOWN_BINARY:
+        elif op in binary_ops:
             JaxRDDLCompiler._check_num_args(expr, 2)                
             lhs, rhs = expr.args
             jax_lhs = self._jax(lhs, info)
             jax_rhs = self._jax(rhs, info)
-            jax_op, jax_param = self._unwrap(self.KNOWN_BINARY[op], expr.id, info)
+            jax_op, jax_param = self._unwrap(binary_ops[op], expr.id, info)
             return self._jax_binary(
                 jax_lhs, jax_rhs, jax_op, jax_param, at_least_int=True)
         
@@ -921,19 +1135,19 @@ class JaxRDDLCompiler:
             f'Control operator {op} is not supported.\n' + 
             print_stack_trace(expr))   
     
-    def _jax_if_helper(self):
-        
-        def _jax_wrapped_if_calc_exact(c, a, b, param):
-            return jnp.where(c, a, b)
-        
-        return _jax_wrapped_if_calc_exact
-    
     def _jax_if(self, expr, info):
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_CAST']
         JaxRDDLCompiler._check_num_args(expr, 3)
-        jax_if, jax_param = self._unwrap(self._jax_if_helper(), expr.id, info)
+        pred, if_true, if_false = expr.args     
+        
+        # if predicate is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(pred):
+            if_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_IF
+        else:
+            if_op = self.IF_HELPER
+        jax_if, jax_param = self._unwrap(if_op, expr.id, info)
         
-        pred, if_true, if_false = expr.args        
+        # recursively compile arguments   
         jax_pred = self._jax(pred, info)
         jax_true = self._jax(if_true, info)
         jax_false = self._jax(if_false, info)
@@ -951,23 +1165,20 @@ class JaxRDDLCompiler:
             
         return _jax_wrapped_if_then_else
     
-    def _jax_switch_helper(self):
-        
-        def _jax_wrapped_switch_calc_exact(pred, cases, param):
-            pred = pred[jnp.newaxis, ...]
-            sample = jnp.take_along_axis(cases, pred, axis=0)
-            assert sample.shape[0] == 1
-            return sample[0, ...]
-
-        return _jax_wrapped_switch_calc_exact
-        
     def _jax_switch(self, expr, info):
-        pred, *_ = expr.args             
+             
+        # if expression is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(expr):
+            switch_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_SWITCH
+        else:
+            switch_op = self.SWITCH_HELPER
+        jax_switch, jax_param = self._unwrap(switch_op, expr.id, info)
+        
+        # recursively compile predicate
+        pred, *_ = expr.args  
         jax_pred = self._jax(pred, info)
-        jax_switch, jax_param = self._unwrap(
-            self._jax_switch_helper(), expr.id, info)
         
-        # wrap cases as JAX expressions
+        # recursively compile cases
         cases, default = self.traced.cached_sim_info(expr) 
         jax_default = None if default is None else self._jax(default, info)
         jax_cases = [(jax_default if _case is None else self._jax(_case, info))
@@ -983,7 +1194,8 @@ class JaxRDDLCompiler:
             for (i, jax_case) in enumerate(jax_cases):
                 sample_cases[i], key, err_case = jax_case(x, params, key)
                 err |= err_case                
-            sample_cases = jnp.asarray(sample_cases)
+            sample_cases = jnp.asarray(
+                sample_cases, dtype=self._fix_dtype(sample_cases))
             
             # predicate (enum) is an integer - use it to extract from case array
             param = params.get(jax_param, None)
@@ -1179,30 +1391,28 @@ class JaxRDDLCompiler:
             scale, key, err2 = jax_scale(x, params, key)
             key, subkey = random.split(key)
             U = random.uniform(key=subkey, shape=jnp.shape(scale), dtype=self.REAL)
-            sample = scale * jnp.power(-jnp.log1p(-U), 1.0 / shape)
+            sample = scale * jnp.power(-jnp.log(U), 1.0 / shape)
             out_of_bounds = jnp.logical_not(jnp.all((shape > 0) & (scale > 0)))
             err = err1 | err2 | (out_of_bounds * ERR)
             return sample, key, err
         
         return _jax_wrapped_distribution_weibull
     
-    def _jax_bernoulli_helper(self):
-        
-        def _jax_wrapped_calc_bernoulli_exact(key, prob, param):
-            return random.bernoulli(key, prob)
-        
-        return _jax_wrapped_calc_bernoulli_exact
-        
     def _jax_bernoulli(self, expr, info):
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_PARAM_BERNOULLI']
         JaxRDDLCompiler._check_num_args(expr, 1)
-        jax_bern, jax_param = self._unwrap(
-            self._jax_bernoulli_helper(), expr.id, info)
-        
         arg_prob, = expr.args
+        
+        # if probability is non-fluent, always use the exact operation
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(arg_prob):
+            bern_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_BERNOULLI
+        else:
+            bern_op = self.BERNOULLI_HELPER
+        jax_bern, jax_param = self._unwrap(bern_op, expr.id, info)
+        
+        # recursively compile arguments
         jax_prob = self._jax(arg_prob, info)
         
-        # uses the implicit JAX subroutine
         def _jax_wrapped_distribution_bernoulli(x, params, key):
             prob, key, err = jax_prob(x, params, key)
             key, subkey = random.split(key)
@@ -1266,8 +1476,8 @@ class JaxRDDLCompiler:
         def _jax_wrapped_distribution_binomial(x, params, key):
             trials, key, err2 = jax_trials(x, params, key)       
             prob, key, err1 = jax_prob(x, params, key)
-            trials = jnp.asarray(trials, self.REAL)
-            prob = jnp.asarray(prob, self.REAL)
+            trials = jnp.asarray(trials, dtype=self.REAL)
+            prob = jnp.asarray(prob, dtype=self.REAL)
             key, subkey = random.split(key)
             dist = tfp.distributions.Binomial(total_count=trials, probs=prob)
             sample = dist.sample(seed=subkey).astype(self.INT)
@@ -1290,11 +1500,10 @@ class JaxRDDLCompiler:
         def _jax_wrapped_distribution_negative_binomial(x, params, key):
             trials, key, err2 = jax_trials(x, params, key)       
             prob, key, err1 = jax_prob(x, params, key)
-            trials = jnp.asarray(trials, self.REAL)
-            prob = jnp.asarray(prob, self.REAL)
+            trials = jnp.asarray(trials, dtype=self.REAL)
+            prob = jnp.asarray(prob, dtype=self.REAL)
             key, subkey = random.split(key)
-            dist = tfp.distributions.NegativeBinomial(
-                total_count=trials, probs=prob)
+            dist = tfp.distributions.NegativeBinomial(total_count=trials, probs=prob)
             sample = dist.sample(seed=subkey).astype(self.INT)
             out_of_bounds = jnp.logical_not(jnp.all(
                 (prob >= 0) & (prob <= 1) & (trials > 0)))
@@ -1316,7 +1525,7 @@ class JaxRDDLCompiler:
             shape, key, err1 = jax_shape(x, params, key)
             rate, key, err2 = jax_rate(x, params, key)
             key, subkey = random.split(key)
-            sample = random.beta(key=subkey, a=shape, b=rate)
+            sample = random.beta(key=subkey, a=shape, b=rate, dtype=self.REAL)
             out_of_bounds = jnp.logical_not(jnp.all((shape > 0) & (rate > 0)))
             err = err1 | err2 | (out_of_bounds * ERR)
             return sample, key, err
@@ -1325,23 +1534,35 @@ class JaxRDDLCompiler:
     
     def _jax_geometric(self, expr, info):
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_PARAM_GEOMETRIC']
-        JaxRDDLCompiler._check_num_args(expr, 1)
-        
+        JaxRDDLCompiler._check_num_args(expr, 1)        
         arg_prob, = expr.args
         jax_prob = self._jax(arg_prob, info)
-        floor_op, jax_param = self._unwrap(
-            self.KNOWN_UNARY['floor'], expr.id, info)
         
-        # reparameterization trick Geom(p) = floor(ln(U(0, 1)) / ln(p)) + 1
-        def _jax_wrapped_distribution_geometric(x, params, key):
-            prob, key, err = jax_prob(x, params, key)
-            key, subkey = random.split(key)
-            U = random.uniform(key=subkey, shape=jnp.shape(prob), dtype=self.REAL)
-            param = params.get(jax_param, None)
-            sample = floor_op(jnp.log1p(-U) / jnp.log1p(-prob), param) + 1
-            out_of_bounds = jnp.logical_not(jnp.all((prob >= 0) & (prob <= 1)))
-            err |= (out_of_bounds * ERR)
-            return sample, key, err
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(arg_prob):            
+            
+            # prob is non-fluent: do not reparameterize
+            def _jax_wrapped_distribution_geometric(x, params, key):
+                prob, key, err = jax_prob(x, params, key)
+                key, subkey = random.split(key)
+                sample = random.geometric(key=subkey, p=prob, dtype=self.INT)
+                out_of_bounds = jnp.logical_not(jnp.all((prob >= 0) & (prob <= 1)))
+                err |= (out_of_bounds * ERR)
+                return sample, key, err
+        
+        else:    
+            floor_op, jax_param = self._unwrap(
+                self.KNOWN_UNARY['floor'], expr.id, info)
+            
+            # reparameterization trick Geom(p) = floor(ln(U(0, 1)) / ln(p)) + 1
+            def _jax_wrapped_distribution_geometric(x, params, key):
+                prob, key, err = jax_prob(x, params, key)
+                key, subkey = random.split(key)
+                U = random.uniform(key=subkey, shape=jnp.shape(prob), dtype=self.REAL)
+                param = params.get(jax_param, None)
+                sample = floor_op(jnp.log(U) / jnp.log(1.0 - prob), param) + 1
+                out_of_bounds = jnp.logical_not(jnp.all((prob >= 0) & (prob <= 1)))
+                err |= (out_of_bounds * ERR)
+                return sample, key, err
         
         return _jax_wrapped_distribution_geometric
     
@@ -1359,7 +1580,7 @@ class JaxRDDLCompiler:
             shape, key, err1 = jax_shape(x, params, key)
             scale, key, err2 = jax_scale(x, params, key)
             key, subkey = random.split(key)
-            sample = scale * random.pareto(key=subkey, b=shape)
+            sample = scale * random.pareto(key=subkey, b=shape, dtype=self.REAL)
             out_of_bounds = jnp.logical_not(jnp.all((shape > 0) & (scale > 0)))
             err = err1 | err2 | (out_of_bounds * ERR)
             return sample, key, err
@@ -1377,7 +1598,8 @@ class JaxRDDLCompiler:
         def _jax_wrapped_distribution_t(x, params, key):
             df, key, err = jax_df(x, params, key)
             key, subkey = random.split(key)
-            sample = random.t(key=subkey, df=df, shape=jnp.shape(df))
+            sample = random.t(
+                key=subkey, df=df, shape=jnp.shape(df), dtype=self.REAL)
             out_of_bounds = jnp.logical_not(jnp.all(df > 0))
             err |= (out_of_bounds * ERR)
             return sample, key, err
@@ -1464,7 +1686,7 @@ class JaxRDDLCompiler:
             scale, key, err2 = jax_scale(x, params, key)
             key, subkey = random.split(key)
             U = random.uniform(key=subkey, shape=jnp.shape(scale), dtype=self.REAL)
-            sample = jnp.log(1.0 - jnp.log1p(-U) / shape) / scale
+            sample = jnp.log(1.0 - jnp.log(U) / shape) / scale
             out_of_bounds = jnp.logical_not(jnp.all((shape > 0) & (scale > 0)))
             err = err1 | err2 | (out_of_bounds * ERR)
             return sample, key, err
@@ -1516,25 +1738,21 @@ class JaxRDDLCompiler:
     # random variables with enum support
     # ===========================================================================
     
-    def _jax_discrete_helper(self):
-        
-        def _jax_wrapped_discrete_calc_exact(key, prob, param):
-            logits = jnp.log(prob)
-            sample = random.categorical(key=key, logits=logits, axis=-1)
-            out_of_bounds = jnp.logical_not(jnp.logical_and(
-                jnp.all(prob >= 0),
-                jnp.allclose(jnp.sum(prob, axis=-1), 1.0)))
-            return sample, out_of_bounds
-        
-        return _jax_wrapped_discrete_calc_exact
-            
     def _jax_discrete(self, expr, info, unnorm):
         NORMAL = JaxRDDLCompiler.ERROR_CODES['NORMAL']
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_PARAM_DISCRETE']
-        jax_discrete, jax_param = self._unwrap(
-            self._jax_discrete_helper(), expr.id, info)
-        
         ordered_args = self.traced.cached_sim_info(expr)
+        
+        # if all probabilities are non-fluent, then always sample exact
+        has_fluent_arg = any(self.traced.cached_is_fluent(arg) 
+                             for arg in ordered_args)
+        if self.compile_non_fluent_exact and not has_fluent_arg:
+            discrete_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_DISCRETE
+        else:
+            discrete_op = self.DISCRETE_HELPER
+        jax_discrete, jax_param = self._unwrap(discrete_op, expr.id, info)
+        
+        # compile probability expressions
         jax_probs = [self._jax(arg, info) for arg in ordered_args]
         
         def _jax_wrapped_distribution_discrete(x, params, key):
@@ -1561,12 +1779,18 @@ class JaxRDDLCompiler:
     
     def _jax_discrete_pvar(self, expr, info, unnorm):
         ERR = JaxRDDLCompiler.ERROR_CODES['INVALID_PARAM_DISCRETE']
-        JaxRDDLCompiler._check_num_args(expr, 1)
-        jax_discrete, jax_param = self._unwrap(
-            self._jax_discrete_helper(), expr.id, info)
-        
+        JaxRDDLCompiler._check_num_args(expr, 2)
         _, args = expr.args
         arg, = args
+        
+        # if probabilities are non-fluent, then always sample exact
+        if self.compile_non_fluent_exact and not self.traced.cached_is_fluent(arg):
+            discrete_op = JaxRDDLCompiler.EXACT_RDDL_TO_JAX_DISCRETE
+        else:
+            discrete_op = self.DISCRETE_HELPER
+        jax_discrete, jax_param = self._unwrap(discrete_op, expr.id, info)
+        
+        # compile probability function
         jax_probs = self._jax(arg, info)
 
         def _jax_wrapped_distribution_discrete_pvar(x, params, key):
@@ -1687,7 +1911,7 @@ class JaxRDDLCompiler:
             out_of_bounds = jnp.logical_not(jnp.all(alpha > 0))
             error |= (out_of_bounds * ERR)
             key, subkey = random.split(key)
-            Gamma = random.gamma(key=subkey, a=alpha)
+            Gamma = random.gamma(key=subkey, a=alpha, dtype=self.REAL)
             sample = Gamma / jnp.sum(Gamma, axis=-1, keepdims=True)
             sample = jnp.moveaxis(sample, source=-1, destination=index)
             return sample, key, error
@@ -1706,8 +1930,8 @@ class JaxRDDLCompiler:
         def _jax_wrapped_distribution_multinomial(x, params, key):
             trials, key, err1 = jax_trials(x, params, key)
             prob, key, err2 = jax_prob(x, params, key)
-            trials = jnp.asarray(trials, self.REAL)
-            prob = jnp.asarray(prob, self.REAL)
+            trials = jnp.asarray(trials, dtype=self.REAL)
+            prob = jnp.asarray(prob, dtype=self.REAL)
             key, subkey = random.split(key)
             dist = tfp.distributions.Multinomial(total_count=trials, probs=prob)
             sample = dist.sample(seed=subkey).astype(self.INT)