final implementation of shc

experiments/stochastic_hillclimber/actors/neuron.py

@@ -17,12 +17,21 @@ class Neuron(Actor):
         # input_idps: [(input_id, [w1, w2, ...])]
         self.inputs = {}
         self.order = []
+
+        """
         for (inp_id, weights) in input_idps:
             self.order.append(inp_id)
             self.inputs[inp_id] = {"weights": list(weights), "got": False, "val": None}
-        # WICHTIG: Bias lernbar machen
-        self.bias = random.uniform(-2.0, 2.0)
-        # Für Backup/Restore
+        """
+
+        self.bias = 0.0
+        for (inp_id, weights) in input_idps:
+            if inp_id == "bias":
+                self.bias = float(weights[0])
+            else:
+                self.order.append(inp_id)
+                self.inputs[inp_id] = {"weights": list(weights), "got": False, "val": None}
+
         self._backup_inputs = None
         self._backup_bias = None
 
@@ -61,13 +70,11 @@ class Neuron(Actor):
                     print(f"Neuron {self.nid}: input_sum={acc + self.bias:.3f}, output={out:.3f}")
 
             elif tag == "get_backup":
-                # Packe Gewichte + Bias zurück
                 idps = [(i, self.inputs[i]["weights"]) for i in self.order]
                 idps.append(("bias", self.bias))
                 await self.cx_pid.send(("backup_from_neuron", self.nid, idps))
 
             elif tag == "weight_backup":
-                # Tiefkopie der Gewichte + Bias
                 print(f"Neuron {self.nid}: backing up weights")
                 self._backup_inputs = {k: {"weights": v["weights"][:]} for k, v in self.inputs.items()}
                 self._backup_bias = self.bias
@@ -79,21 +86,17 @@ class Neuron(Actor):
                     self.bias = self._backup_bias
 
             elif tag == "weight_perturb":
-                # In neuron.py weight_perturb, add:
                 print(f"Neuron {self.nid}: perturbing {len([w for i in self.order for w in self.inputs[i]['weights']])} weights")
-                # MP wie im Buch: 1/sqrt(TotalWeightsIncludingBias)
                 tot_w = sum(len(self.inputs[i]["weights"]) for i in self.order) + 1
                 mp = 1 / math.sqrt(tot_w)
                 delta_mag = 2.0 * math.pi
                 sat_lim = 2.0 * math.pi
 
-                # Gewichte perturbieren
                 for i in self.order:
                     ws = self.inputs[i]["weights"]
                     for j in range(len(ws)):
                         if random.random() < mp:
                             ws[j] = _sat(ws[j] + (random.random() - 0.5) * delta_mag, -sat_lim, sat_lim)
-                # Bias perturbieren
                 if random.random() < mp:
                     self.bias = _sat(self.bias + (random.random() - 0.5) * delta_mag, -sat_lim, sat_lim)