Merge branch 'main' into alizaidi/exog_var_lookup

ddm-cartpole.ink

@@ -0,0 +1,87 @@
+# https://docs.microsoft.com/en-us/bonsai/inkling/basics
+inkling "2.0"
+using Math
+using Goal
+
+# The maximum angle (in radians) before fallover or failure
+const MaxPoleAngle = (12 * 2 * Math.Pi) / 360
+# Length of the cartpole track in meters
+const TrackLength = 2.0
+const MaxPosition = TrackLength / 2
+
+
+# These are all the values the sim produces
+type SimState {
+    cart_position: number,  # (m). 0 is the center of the track
+    cart_velocity: number,  # (m/s)
+    pole_angle: number,  # (rad). 0 is vertical.
+    pole_angular_velocity: number,  # (rad/s)
+    # pole_center_position: number,  # (m) -- x position of the center of the pole 
+    # pole_center_velocity: number,  # (m/s) -- x velocity of the center of the pole
+    # target_pole_position: number,  # (m)
+
+    # physical params from the config are available too
+    # cart_mass: number,  # (kg)
+    # pole_mass: number,  # (kg)
+    # pole_length: number,  # (m)
+}
+
+# This is a subset of the SimState that we'll make available to the brain
+# (these should all be values that will be available to a deployed brain)
+type ObservedState {
+    cart_position: number,  # (m). 0 is the center of the track
+    cart_velocity: number,  # (m/s)
+    pole_angle: number,  # (rad). 0 is vertical.
+    pole_angular_velocity: number,  # (rad/s)
+}
+
+type Action {
+    # Force to apply, up to the max force magnitude (1 Newton by default)
+    command: number<-1..1>
+}
+
+# Configuration variables for the simulator
+type SimConfig {
+    cart_mass: number,  # (kg), default 0.31
+    pole_mass: number,  # (kg), default 0.055
+    pole_length: number,  # (m), default 0.4
+    initial_cart_position: number<-MaxPosition .. MaxPosition>,  # (m), default 0 (center)
+    initial_cart_velocity: number,   # (m/s), default 0
+    initial_pole_angle: number,  # (rad), default 0
+    initial_angular_velocity: number,  # (rad/s), default 0
+    # target_pole_position: number<-MaxPosition .. MaxPosition>,  # (m), default 0
+}
+
+graph (input: ObservedState): Action {
+    concept BalancePole(input): Action {
+        curriculum {
+            source simulator (Action: Action, Config: SimConfig): SimState {
+                package "ddm-cartpole-08-31"
+            }
+            training {
+                EpisodeIterationLimit: 200,
+            }
+            goal (SimState: SimState) {
+                avoid FallOver:
+                    Math.Abs(SimState.pole_angle) in Goal.RangeAbove(MaxPoleAngle)
+                avoid OutOfRange:
+                    Math.Abs(SimState.cart_position) in Goal.RangeAbove(MaxPosition)
+            }
+
+            lesson initial_positions  {
+                scenario {
+                    initial_cart_position: number<-0.1..0.1>
+                }
+            }
+            lesson all_positions {
+                scenario {
+                    initial_cart_position: number<-MaxPosition .. MaxPosition>
+                }
+            }
+        }
+    }
+}
+
+# Special string to hook up the simulator visualizer
+# in the web interface.
+const SimulatorVisualizer = "/cartpoleviz/"

requirements-tests.txt

@@ -3,9 +3,9 @@ hyperopt==0.2.5
 lightgbm==3.1.0
 matplotlib==3.4.1
 microsoft-bonsai-api==0.1.3
-mlflow==1.15.0
+mlflow==1.23.1
 natsort==7.1.0
-numpy==1.19.5
+numpy==1.22.0
 optuna==2.6.0
 pandas==1.2.4
 python-dotenv==0.13.0

requirements.txt

@@ -24,4 +24,5 @@ scikit-learn==1.1.3
 # skorch==0.12.1
 # tensorboard==2.11.0
 # tune-sklearn==0.4.5
-xgboost==1.7.1
+xgboost==1.7.1
+

torch_models.py

@@ -84,7 +84,9 @@ class PyTorchModel(BaseModel):
         else:
             use_cuda = torch.cuda.is_available()
             self.device = torch.device("cuda" if use_cuda else "cpu")
-
+        
+        # For more information about this class configs, you can see the following link:
+        # https://skorch.readthedocs.io/en/stable/regressor.html
         self.model = NeuralNetRegressor(
             network,
             device=self.device,