Latest changes for error and perf checking

ch-13/4.3-monte-carlo_perf/display_arena.py

@@ -0,0 +1,16 @@
+from matplotlib import pyplot as plt
+
+from robot import arena
+
+print(arena.distance_grid.min(), arena.distance_grid.max())
+for line in arena.boundary_lines:
+    plt.plot([line[0][0], line[1][0]], [line[0][1], line[1][1]], color="black")
+overscan_size = arena.overscan * arena.grid_cell_size
+plt.imshow(
+    arena.distance_grid.T,
+    extent = [-overscan_size, arena.width + overscan_size, -overscan_size, arena.height + overscan_size], 
+    origin="lower",
+    cmap="gray"
+)
+
+plt.show()

ch-13/4.3-monte-carlo_perf/robot/arena.py

@@ -9,6 +9,8 @@ height = 1500
 cutout_width = 500
 cutout_height = 500
 
+low_probability = 10 ** -10
+
 boundary_lines = [
     [(0,0), (0, height)],
     [(0, height), (width, height)],
@@ -65,7 +67,7 @@ def get_distance_likelihood(x, y):
         distance = get_distance_to_segment(x, y, segment)
         if min_distance is None or distance < min_distance:
             min_distance = distance
-    return 1.0 / (1 + min_distance/250) ** 2
+    return 1.0 / (1 + min_distance/100) ** 2
 
 
 # beam endpoint model
@@ -88,10 +90,10 @@ def make_distance_grid():
 
 distance_grid = make_distance_grid()
 
-def get_distance_grid_at_point(x, y):
+def get_distance_likelihood_at(x, y):
   """Return the distance grid value at the given point."""
   grid_x = int(x // grid_cell_size + overscan)
   grid_y = int(y // grid_cell_size + overscan)
   if grid_x < 0 or grid_x >= distance_grid.shape[0] or grid_y < 0 or grid_y >= distance_grid.shape[1]:
-    return 0
+    return low_probability
   return distance_grid[grid_x, grid_y]

ch-13/4.3-monte-carlo_perf/robot/code.py

@@ -79,16 +79,16 @@ class Simulation:
                 int(random.uniform(0, arena.width)),
                 int(random.uniform(0, arena.height)),
                 int(random.uniform(0, 360))) for _ in range(self.population_size)],
-            dtype=np.float,
+            dtype=np.int16,
         )
         self.distance_sensors = DistanceSensorTracker()
         self.collision_avoider = CollisionAvoid(self.distance_sensors)
         self.last_encoder_left = robot.left_encoder.read()
         self.last_encoder_right = robot.right_encoder.read()
-        self.alpha_rot = 0.05
+        self.alpha_rot = 0.09
-        self.alpha_rot_trans = 0.01
+        self.alpha_rot_trans = 0.05
-        self.alpha_trans = 0.05
+        self.alpha_trans = 0.12
-        self.alpha_trans_rot = 0.01
+        self.alpha_trans_rot = 0.05
         
         # profiling
         self.pc_resample = PerformanceCounter()
@@ -99,26 +99,6 @@ class Simulation:
         self.pc_observe_distance_sensors = PerformanceCounter()
         self.pc_observation_model = PerformanceCounter()
 
-    def resample(self, weights, sample_count):
-        """Return sample_count number of samples from the
-        poses, based on the weights array.
-        Uses low variance resampling"""
-        self.pc_resample.start()
-        samples = np.zeros((sample_count, 3))
-        interval = 1 / sample_count
-        shift = random.uniform(0, interval)
-        cumulative_weights = weights[0]
-        source_index = 0
-        for current_index in range(sample_count):
-            weight_index = shift + current_index * interval
-            while weight_index >= cumulative_weights:
-                source_index += 1
-                source_index = min(len(weights), source_index)
-                cumulative_weights += weights[source_index]
-            samples[current_index] = self.poses[source_index]
-        self.pc_resample.stop()
-        return samples
-
     def convert_odometry_to_motion(self, left_encoder_delta, right_encoder_delta):
         """
         left_encoder is the change in the left encoder
@@ -203,12 +183,14 @@ class Simulation:
         right_hypotenuse = np.sqrt(opposite**2 + adjacent**2)
 
         # modify the current weights based on the distance sensors
-        left_sensor = np.zeros((self.poses.shape[0], 2), dtype=np.float)
+        # left_sensor = np.zeros((self.poses.shape[0], 2), dtype=np.float)
         right_sensor = np.zeros((self.poses.shape[0], 2), dtype=np.float)
         # left sensor
         poses_left_angle = np.radians(self.poses[:, 2]) + left_angle
-        left_sensor[:, 0] = self.poses[:, 0] + np.cos(poses_left_angle) * left_hypotenuse
+        left_sensor = np.concatenate([
-        left_sensor[:, 1] = self.poses[:, 1] + np.sin(poses_left_angle) * left_hypotenuse
+            self.poses[:, 0] + np.cos(poses_left_angle) * left_hypotenuse,
+            self.poses[:, 1] + np.sin(poses_left_angle) * left_hypotenuse
+        ], axis=1)
 
         # right sensor
         poses_right_angle = np.radians(self.poses[:, 2]) - right_angle
@@ -217,8 +199,8 @@ class Simulation:
 
         # Look up the distance in the arena
         for index in range(self.poses.shape[0]):
-            sensor_weight = arena.get_distance_grid_at_point(left_sensor[index,0], left_sensor[index,1])
+            sensor_weight = arena.get_distance_likelihood_at(left_sensor[index,0], left_sensor[index,1])
-            sensor_weight += arena.get_distance_grid_at_point(right_sensor[index,0], right_sensor[index,1])
+            sensor_weight += arena.get_distance_likelihood_at(right_sensor[index,0], right_sensor[index,1])
             weights[index] *= sensor_weight
         self.pc_observe_distance_sensors.stop()
         return weights
@@ -228,12 +210,38 @@ class Simulation:
         weights = np.ones(self.poses.shape[0], dtype=np.float)
         for index, pose in enumerate(self.poses):
             if not arena.contains(pose[:1], pose[:2]):
-                weights[index] = 0.01
+                weights[index] = arena.low_probability
         weights = self.observe_distance_sensors(weights)
-        weights = weights / np.sum(weights)
         self.pc_observation_model.stop()
         return weights
 
+    def resample(self, weights, sample_count):
+        """Return sample_count number of samples from the
+        poses, based on the weights array.
+        Uses low variance resampling"""
+        self.pc_resample.start()
+        samples = np.zeros((sample_count, 3))
+        interval = np.sum(weights) / sample_count
+        shift = random.uniform(0, interval)
+        cumulative_weights = weights[0]
+        source_index = 0
+        try:
+            for current_index in range(sample_count):
+                weight_index = shift + current_index * interval
+                while weight_index >= cumulative_weights:
+                    source_index += 1
+                    source_index = min(len(weights), source_index)
+                    cumulative_weights += weights[source_index]
+                samples[current_index] = self.poses[source_index]
+        except IndexError:
+            send_json({"error": "IndexError in resample.", "weights": [weights.tolist()]})
+            raise
+        if samples.shape[0] != sample_count:
+            send_json({"error": "Sample count mismatch in resample.", "samples": [samples.tolist()]})
+            raise Exception("Sample count mismatch in resample.")
+        self.pc_resample.stop()
+        return samples
+
     def print_pc_lines(self):
         if self.pc_odometry.count % 10 != 0:
             return

ch-13/4.3-monte-carlo_perf/robot/performance_counter.py

@@ -16,6 +16,8 @@ class PerformanceCounter:
         self.count += 1
 
     def per_call(self):
+        if self.count == 0:
+            return 0
         return self.total_time / self.count
     
     def total_call_time(self):