fix: scaling pdfs with action max/min (#23)

* fix: scaling pdfs with action max/min

* test: additional test of TreeParameters.construct

* fix: scaling the probabilities inside the update function

README.md

@@ -34,7 +34,7 @@ and much more on CATX and its inner workings.
   author = {Wissam Bejjani and Cyprien Courtot},
   title = {CATX: contextual bandits library for Continuous Action Trees with smoothing in JAX},
   url = {https://github.com/instadeepai/catx/},
-  version = {0.2.0},
+  version = {0.2.1},
   year = {2022},
 }
 ```

catx/VERSION

@@ -1 +1 @@
-0.2.0
+0.2.1

catx/catx.py

@@ -348,6 +348,7 @@ def _forward(
 
             # Scale sampled actions to the environment action range.
             actions = actions * (self._action_max - self._action_min) + self._action_min
+            probabilities /= self._action_max - self._action_min
 
             return actions, probabilities
 
@@ -553,6 +554,7 @@ def _update(
 
         # Scale actions from the environment action range to the tree action range.
         actions = (actions - self._action_min) / (self._action_max - self._action_min)
+        probabilities *= self._action_max - self._action_min
 
         smooth_costs = self._compute_smooth_costs(
             costs=costs, actions=actions, probabilities=probabilities

tests/test_tree.py

@@ -81,7 +81,7 @@ def _forward(
             tree_params=tree_parameters,
         )
 
-        output_logits = tree(obs=x, network_extras=network_extras)
+        output_logits: Dict[int, Logits] = tree(obs=x, network_extras=network_extras)
 
         # Validate the tree has as many neural networks as depth.
         assert jnp.shape(jax.tree_leaves(tree.networks))[0] == tree_parameters.depth
@@ -116,3 +116,13 @@ def _forward(
         assert logits_shape[d] == (jnp.shape(jax_observations)[0], 2**d, 2)
 
     chex.assert_tree_all_finite(logits)
+
+
+def test_tree_parameters__probabilities_and_volumes() -> None:
+    tree_param = TreeParameters.construct(bandwidth=1 / 4, discretization_parameter=4)
+
+    expected_volumes = jnp.full_like(tree_param.volumes, 1 / 2)
+    expected_probabilities = jnp.full_like(tree_param.probabilities, 2)
+
+    assert jnp.allclose(tree_param.volumes, expected_volumes)
+    assert jnp.allclose(tree_param.probabilities, expected_probabilities)