woche12 > master: code py - rohe implementierung der walks

code/python/models/commands-schema.yaml

@@ -170,9 +170,21 @@ components:
           $ref: '#/components/schemas/DataTypeLandscapeGeometry'
         optimise:
           $ref: '#/components/schemas/EnumOptimiseMode'
+        coords-init:
+          description: Initial co-ordinates to start the algorithm.
+          type: array
+          items:
+            type: integer
+          minItems: 1
+        temperature-init:
+          type: float
+          default: 1.
         annealing:
           type: boolean
           default: false
+        one-based:
+          type: boolean
+          default: false
     # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     # Algorithm: Genetic Algorithm
     # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -243,10 +255,16 @@ components:
       type: object
       required:
         - neighbourhoods
+        - labels
         - values
       properties:
         neighbourhoods:
           $ref: '#/components/schemas/DataTypeLandscapeNeighbourhoods'
+        labels:
+          type: array
+          items:
+            type: string
+          minItems: 1
         values:
           $ref: '#/components/schemas/DataTypeLandscapeValues'
     DataTypeLandscapeNeighbourhoods:

code/python/src/algorithms/random_walk/algorithms.py

@@ -32,14 +32,50 @@ __all__ = [
 def adaptive_walk_algorithm(
     landscape: Landscape,
     r: float,
+    coords_init: tuple,
     optimise: EnumOptimiseMode,
     verbose: bool,
 ):
     '''
     Führt den Adapative-Walk Algorithmus aus, um ein lokales Minimum zu bestimmen.
     '''
-    log_warn('Noch nicht implementiert!');
-    return;
+
+    # lege Fitness- und Umgebungsfunktionen fest:
+    match optimise:
+        case EnumOptimiseMode.max:
+            f = lambda x: -landscape.fitness(*x);
+        case _:
+            f = lambda x: landscape.fitness(*x);
+    nbhd = lambda x: landscape.neighbourhood(*x, r=r, strict=True);
+    label = lambda x: landscape.label(*x);
+
+    # initialisiere
+    x = coords_init;
+    fx = f(x);
+    fy = fx;
+    N = nbhd(x);
+
+    # führe walk aus:
+    while True:
+        # Wähle zufälligen Punkt und berechne fitness-Wert:
+        y = uniform_random_choice(N);
+        fy = f(y);
+
+        # Nur dann aktualisieren, wenn sich f-Wert verbessert:
+        if fy < fx:
+            # Punkt + Umgebung + f-Wert aktualisieren
+            x = y;
+            fx = fy;
+            N = nbhd(x);
+        else:
+            # Nichts machen!
+            pass;
+
+        # Nur dann (erfolgreich) abbrechen, wenn f-Wert lokal Min:
+        if fx <= min([f(y) for y in N], default=fx):
+            break;
+
+    return x;
 
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 # METHOD gradient walk
@@ -48,14 +84,56 @@ def adaptive_walk_algorithm(
 def gradient_walk_algorithm(
     landscape: Landscape,
     r: float,
+    coords_init: tuple,
     optimise: EnumOptimiseMode,
     verbose: bool,
 ):
     '''
     Führt den Gradient-Descent (bzw. Ascent) Algorithmus aus, um ein lokales Minimum zu bestimmen.
     '''
-    log_warn('Noch nicht implementiert!');
-    return;
+
+    # lege Fitness- und Umgebungsfunktionen fest:
+    match optimise:
+        case EnumOptimiseMode.max:
+            f = lambda x: -landscape.fitness(*x);
+        case _:
+            f = lambda x: landscape.fitness(*x);
+    nbhd = lambda x: landscape.neighbourhood(*x, r=r, strict=True);
+    label = lambda x: landscape.label(*x);
+
+    # initialisiere
+    x = coords_init;
+    fx = landscape.fitness(*x);
+    fy = fx;
+    N = nbhd(x);
+    f_values = [f(y) for y in N];
+    fmin = min(f_values);
+    Z = [y for y, fy in zip(N, f_values) if fy == fmin];
+
+    # führe walk aus:
+    while True:
+        # Wähle zufälligen Punkt mit steilstem Abstieg und berechne fitness-Wert:
+        y = uniform_random_choice(Z);
+        fy = fmin;
+
+        # Nur dann aktualisieren, wenn sich f-Wert verbessert:
+        if fy < fx:
+            # Punkt + Umgebung + f-Wert aktualisieren
+            x = y;
+            fx = fy;
+            N = nbhd(y);
+            f_values = [f(y) for y in N];
+            fmin = min(f_values);
+            Z = [y for y, fy in zip(N, f_values) if fy == fmin];
+        else:
+            # Nichts machen!
+            pass;
+
+        # Nur dann (erfolgreich) abbrechen, wenn f-Wert lokal Min:
+        if fx <= min([f(y) for y in N], default=fx):
+            break;
+
+    return x;
 
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 # METHOD metropolis walk
@@ -64,6 +142,8 @@ def gradient_walk_algorithm(
 def metropolis_walk_algorithm(
     landscape: Landscape,
     r: float,
+    coords_init: tuple,
+    T: float,
     annealing: bool,
     optimise: EnumOptimiseMode,
     verbose: bool,
@@ -71,5 +151,56 @@ def metropolis_walk_algorithm(
     '''
     Führt den Metropolis-Walk Algorithmus aus, um ein lokales Minimum zu bestimmen.
     '''
-    log_warn('Noch nicht implementiert!');
-    return;
+
+    # lege Fitness- und Umgebungsfunktionen fest:
+    match optimise:
+        case EnumOptimiseMode.max:
+            f = lambda x: -landscape.fitness(*x);
+        case _:
+            f = lambda x: landscape.fitness(*x);
+    nbhd = lambda x: landscape.neighbourhood(*x, r=r, strict=True);
+    label = lambda x: landscape.label(*x);
+
+    # initialisiere
+    x = coords_init;
+    fx = f(x);
+    fy = fx;
+    nbhd_x = nbhd(x);
+
+    # führe walk aus:
+    k = 0;
+    while True:
+        # Wähle zufälligen Punkt und berechne fitness-Wert:
+        y = uniform_random_choice(nbhd_x);
+        r = uniform(0,1);
+        fy = f(y);
+
+        # Nur dann aktualisieren, wenn sich f-Wert verbessert:
+        if fy < fx or r < math.exp(-(fy-fx)/T):
+            # Punkt + Umgebung + f-Wert aktualisieren
+            x = y;
+            fx = fy;
+            nbhd_x = nbhd(x);
+        else:
+            # Nichts machen!
+            pass;
+
+        # »Temperatur« ggf. abkühlen:
+        if annealing:
+            T = cool_temperature(T, k);
+
+        # Nur dann (erfolgreich) abbrechen, wenn f-Wert lokal Min:
+        if fx <= min([f(y) for y in nbhd_x], default=fx):
+            break;
+
+        k += 1;
+
+    return x;
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# AUXILIARY METHODS
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+def cool_temperature(T: float, k: int, const: float = 1.) -> float:
+    harm = const*(k + 1);
+    return T/(1 + T/harm);

code/python/src/endpoints/ep_algorithm_random_walk.py

@@ -27,15 +27,28 @@ __all__ = [
 
 @run_safely()
 def endpoint_random_walk(command: CommandRandomWalk) -> Result[CallResult, CallError]:
+    # Compute landscape (fitness fct + topology) + initial co-ordinates:
+    one_based = command.one_based;
     landscape = Landscape(
         values = command.landscape.values,
+        labels = command.landscape.labels,
         metric = command.landscape.neighbourhoods.metric,
+        one_based = one_based,
     );
+    if isinstance(command.coords_init, list):
+        coords_init = tuple(command.coords_init);
+        if one_based:
+            coords_init = tuple(xx - 1 for xx in coords_init);
+        assert len(coords_init) == landscape.dim, 'Dimension of initial co-ordinations inconsistent with landscape!';
+    else:
+        coords_init = landscape.coords_middle;
+
     match command.algorithm:
         case EnumWalkMode.adaptive:
             result = adaptive_walk_algorithm(
                 landscape = landscape,
                 r = command.landscape.neighbourhoods.radius,
+                coords_init = coords_init,
                 optimise = command.optimise,
                 verbose = config.OPTIONS.random_walk.verbose
             );
@@ -43,6 +56,7 @@ def endpoint_random_walk(command: CommandRandomWalk) -> Result[CallResult, CallE
             result = gradient_walk_algorithm(
                 landscape = landscape,
                 r = command.landscape.neighbourhoods.radius,
+                coords_init = coords_init,
                 optimise = command.optimise,
                 verbose = config.OPTIONS.random_walk.verbose
             );
@@ -50,6 +64,8 @@ def endpoint_random_walk(command: CommandRandomWalk) -> Result[CallResult, CallE
             result = metropolis_walk_algorithm(
                 landscape = landscape,
                 r = command.landscape.neighbourhoods.radius,
+                coords_init = coords_init,
+                T = command.temperature_init,
                 annealing = command.annealing,
                 optimise = command.optimise,
                 verbose = config.OPTIONS.random_walk.verbose

code/python/src/models/random_walk/landscape.py

@@ -27,16 +27,23 @@ __all__ = [
 
 class Landscape():
     _fct: np.ndarray;
+    _labels: list[str];
     _metric: EnumLandscapeMetric;
     _radius: float;
+    _one_based: bool;
 
     def __init__(
         self,
         values: DataTypeLandscapeValues,
+        labels: List[str],
         metric: EnumLandscapeMetric = EnumLandscapeMetric.maximum,
+        one_based: bool = False,
     ):
         self._fct = convert_to_nparray(values);
+        assert len(labels) == self.dim, 'A label is required for each axis/dimension!';
+        self._labels = labels;
         self._metric = metric;
+        self._one_based = one_based;
         return;
 
     @property
@@ -47,14 +54,27 @@ class Landscape():
     def dim(self) -> int:
         return len(self._fct.shape);
 
+    @property
+    def coords_middle(self) -> tuple:
+        return tuple(math.floor(s/2) for s in self.shape);
+
     def fitness(self, *x: int) -> float:
         return self._fct[x];
 
+    def label(self, *x: int) -> str:
+        if self._one_based:
+            x = tuple(xx + 1 for xx in x);
+        expr = ','.join([ f'{name}{xx}' for name, xx in zip(self._labels, x)]);
+        if self.dim > 1:
+            expr = f'({expr})';
+        return expr;
+
     def neighbourhood(self, *x: int, r: float, strict: bool = False) -> List[tuple]:
+        r = int(r);
         sides = [
-            [ xx - j for j in range(1,r+1) if xx - j in range(s) ]
+            [ xx - j for j in range(1, r+1) if xx - j in range(s) ]
             + ([ xx ] if xx in range(s) else [])
-            + [ xx + j for j in range(1,r+1) if xx + j in range(s) ]
+            + [ xx + j for j in range(1, r+1) if xx + j in range(s) ]
             for xx, s in zip(x, self.shape)
         ];
         match self._metric:

code/python/src/thirdparty/maths.py

@@ -10,6 +10,8 @@ import math;
 import numpy as np;
 import pandas as pd;
 import random;
+from random import uniform;
+from random import choice as uniform_random_choice;
 from tabulate import tabulate;
 
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -22,5 +24,7 @@ __all__ = [
     'np',
     'pd',
     'random',
+    'uniform',
+    'uniform_random_choice',
     'tabulate',
 ];