#!/usr/bin/env python3 # coding: utf-8 # # Evolutionary Algorithm # the Evolutionary Algorithm that is supposed to create new recipes based on the Recipe Matrices that are created during the *Recipe Analysis* step. # # The Population of the Evolutional Algorithm consists of a set of recipe trees. Each Recipe Tree consists of several Nodes where each node is of one of the following Types: # # * **Ingredient Node:** # these are the leaf nodes. Containing an ingredient. The score is determined by the actions, that are applied if you follow up the path. At the Moment it measures how many duplicate actions are applied. # * **Action Node:** # An Action that is applied on it's child and this child's subtree. Score indicates the average likelihood that this action is applied on the ingredients inside the subtree # * **Mix Node:** # Mixing ingredients together. This is also the only Node that can have more than one child. The score is the average of all pairwise likelihoods that two ingredients are mixed togethter import sys sys.path.append("../") sys.path.append("../RecipeAnalysis/") import settings import pycrfsuite import json import db.db_settings as db_settings from db.database_connection import DatabaseConnection from Tagging.conllu_generator import ConlluGenerator from Tagging.crf_data_generator import * from RecipeAnalysis.Recipe import Ingredient import ea_tools from difflib import SequenceMatcher import numpy as np import ActionGroups as AG import plotly.graph_objs as go from plotly.offline import download_plotlyjs, init_notebook_mode, plot, iplot from plotly.subplots import make_subplots init_notebook_mode(connected=True) from graphviz import Digraph import itertools import random import plotly.io as pio pio.renderers.default = "jupyterlab" from IPython.display import Markdown, HTML, display from tqdm.autonotebook import tqdm from copy import deepcopy def gaussian(x, mu, sig): return 1./(np.sqrt(2.*np.pi)*sig)*np.exp(-np.power((x - mu)/sig, 2.)/2) # ## load adjacency matrices import dill m_act = dill.load(open("../RecipeAnalysis/m_act.dill", "rb")) m_mix = dill.load(open("../RecipeAnalysis/m_mix.dill", "rb")) m_base_act = dill.load(open("../RecipeAnalysis/m_base_act.dill", "rb")) m_base_mix = dill.load(open("../RecipeAnalysis/m_base_mix.dill", "rb")) m_grouped_mix = dill.load(open("../RecipeAnalysis/m_grouped_mix_raw.dill", "rb")) m_grouped_act = dill.load(open("../RecipeAnalysis/m_grouped_act_raw.dill", "rb")) m_grouped_base_act = dill.load(open("../RecipeAnalysis/m_grouped_base_act_raw.dill", "rb")) #m_act.apply_threshold(3) #m_mix.apply_threshold(3) #m_base_act.apply_threshold(5) #m_base_mix.apply_threshold(5) #c_act = m_act.get_csr() #c_mix = m_mix.get_csr() #c_base_act = m_base_act.get_csr() #c_base_mix = m_base_mix.get_csr() m_act.compile() m_mix.compile() m_base_act.compile() m_base_mix.compile() m_grouped_mix.compile() m_grouped_act.compile() m_grouped_base_act.compile() c_act = m_act._csr c_mix = m_mix._csr c_base_act = m_base_act._csr c_base_mix = m_base_mix._csr c_grouped_mix = m_grouped_mix._csr c_grouped_act = m_grouped_act._csr c_grouped_base_act = m_grouped_base_act._csr actions = m_act.get_labels()[0] base_ingredients = m_base_mix.get_labels() sym_label_buffer = {} fw_label_buffer = {} bw_label_buffer = {} # ### helper functions for adjacency matrices def get_sym_adjacent(key, m, c): index = m._label_index[key] i1 = c[index,:].nonzero()[1] i2 = c[:,index].nonzero()[0] i = np.concatenate((i1,i2)) if m in sym_label_buffer: names = sym_label_buffer[m][i] else: names = np.array(m.get_labels()) sym_label_buffer[m] = names names = names[i] counts = np.concatenate((c[index, i1].toarray().flatten(), c[i2, index].toarray().flatten())) s = np.argsort(-counts) return names[s], counts[s] def get_forward_adjacent(key, m, c): index = m._x_label_index[key] i = c[index,:].nonzero()[1] if m in fw_label_buffer: names = fw_label_buffer[m][i] else: names = np.array(m._y_labels) fw_label_buffer[m] = names names = names[i] counts = c[index, i].toarray().flatten() s = np.argsort(-counts) return names[s], counts[s] def get_backward_adjacent(key, m, c): index = m._y_label_index[key] i = c[:,index].nonzero()[0] if m in bw_label_buffer: names = bw_label_buffer[m][i] else: names = np.array(m._x_labels) bw_label_buffer[m] = names names = names[i] counts = c[i, index].toarray().flatten() s = np.argsort(-counts) return names[s], counts[s] def sym_sum(key, m, c): return np.sum(get_sym_adjacent(key,m,c)[1]) def fw_sum(key, m, c): return np.sum(get_forward_adjacent(key,m,c)[1]) def bw_sum(key, m, c): return np.sum(get_backward_adjacent(key,m,c)[1]) def to_grouped_ingredient(ing:Ingredient): groups = set() for act in ing._action_set: groups.add(AG.groups[act]) grouped_ingredient = Ingredient(ing._base_ingredient) for g in groups: grouped_ingredient.apply_action(g) return grouped_ingredient # #### different score functions # ##### normalizations def fw_normalization_factor(key, m, c, quotient_func): ia = m._x_label_index[key] occurances = c[ia,:].nonzero()[1] return 1. / quotient_func(c[ia,occurances].toarray()) def bw_normalization_factor(key, m, c, quotient_func): ib = m._y_label_index[key] occurances = c[:,ib].nonzero()[0] return 1. / quotient_func(c[occurances,ib].toarray()) def sym_normalization_factor(key, m, c, quotient_func): ii = m._label_index[key] fw_occurances = c[ii,:].nonzero()[1] bw_occurances = c[:,ii].nonzero()[0] return 1. / quotient_func(np.concatenate( [c[ii,fw_occurances].toarray().flatten(), c[bw_occurances,ii].toarray().flatten()] )) def sym_p_a_given_b(key_a, key_b, m, c, quot_func = np.max): ia = m._label_index[key_a] ib = m._label_index[key_b] v = c[ia,ib] + c[ib,ia] return v * sym_normalization_factor(key_b, m, c, quot_func) def fw_p_a_given_b(key_a, key_b, m, c, quot_func = np.max): ia = m._x_label_index[key_a] ib = m._y_label_index[key_b] v = c[ia,ib] return v * bw_normalization_factor(key_b, m, c, quot_func) def bw_p_a_given_b(key_a, key_b, m, c, quot_func = np.max): ia = m._y_label_index[key_a] ib = m._x_label_index[key_b] v = c[ib,ia] return v * fw_normalization_factor(key_b, m, c, quot_func) def sym_score(key_a, key_b, m, c): ia = m._label_index[key_a] ib = m._label_index[key_b] v = c[ia,ib] + c[ib,ia] if v == 0: return 0 return max((v/sym_sum(key_a, m, c)), (v/sym_sum(key_b, m, c))) def asym_score(key_a, key_b, m, c): ia = m._x_label_index[key_a] ib = m._y_label_index[key_b] v = c[ia,ib] if v == 0: return 0 return max(v/fw_sum(key_a, m, c), v/bw_sum(key_b, m, c)) def p_ingredient_unprepared(base_ing): ing = Ingredient(base_ing) base_sum = sym_sum(base_ing, m_base_mix, c_base_mix) specialized_sum = sym_sum(ing.to_json(), m_mix, c_mix) return specialized_sum / base_sum # **new probability for preprocess ingredients:** PREPARE_RATIO_THRESHOLD = 0.35 HEAT_RATIO_THRESHOLD = 0.65 PREPARE_SCORE_EPS = 0.1 HEAT_SCORE_EPS = 0.1 def prepare_ratio(ing:str): try: keys, values = m_grouped_act.get_backward_adjacent(Ingredient(ing).to_json()) except KeyError: return 0 action_dict = dict(zip(keys,values)) if 'prepare' not in action_dict: return 0 if 'heat' not in action_dict: return 1 return action_dict['prepare'] / action_dict['heat'] def random_prepare(ing:str): """ returns randomly a boolean value if ing should be prepared, w.r.t. the prepare_ration function """ return prepare_ratio(ing) > np.random.normal(PREPARE_RATIO_THRESHOLD,0.1) def heat_ratio(ingredient:str): try: action_set, action_weights = m_grouped_base_act.get_backward_adjacent(ingredient) except KeyError: return 0 d = dict(zip(action_set, action_weights)) if 'prepare' not in d: return 1 if 'heat' not in d: return 0 ratio = 1 - d['prepare'] / d['heat'] return ratio def random_heated(ingredient:str): ratio = heat_ratio(ingredient) return ratio > np.random.normal(HEAT_RATIO_THRESHOLD,0.15) def prepare_score(ingredient:Ingredient): ing_str = ingredient._base_ingredient g_ing = to_grouped_ingredient(ingredient) ratio = prepare_ratio(ing_str) if ratio > PREPARE_RATIO_THRESHOLD + PREPARE_SCORE_EPS: if 'prepare' not in g_ing._action_set: return 0 if ratio < PREPARE_RATIO_THRESHOLD - PREPARE_SCORE_EPS: if 'prepare' in g_ing._action_set: return 0 return 1 def heat_score(ingredient:Ingredient): ing_str = ingredient._base_ingredient g_ing = to_grouped_ingredient(ingredient) ratio = heat_ratio(ing_str) if ratio > HEAT_RATIO_THRESHOLD + HEAT_SCORE_EPS: if 'heat' not in g_ing._action_set: return 0 if ratio < HEAT_RATIO_THRESHOLD - HEAT_SCORE_EPS: if 'heat' in g_ing._action_set: return 0 return 1 def relative_action_rank(ingredient:str, action:str): action_set, action_weights = m_base_act.get_backward_adjacent(ingredient) if action not in action_set or len(action_set) <= 1: return 0 return 1 - action_set.tolist().index(action) / (len(action_set) - 1) def filter_set_by_group(act_set, act_w, group): new_act_set = [] new_act_w = [] for i in range(len(act_set)): if act_set[i] in AG.inverse_groups[group]: new_act_set.append(act_set[i]) new_act_w.append(act_w[i]) return np.array(new_act_set), np.array(new_act_w) # ## better normalized scores: def normalized_score(key, matrix): sum_key = matrix.get_sum(key) keys, values = matrix.get_adjacent(key) normalized_values = np.array([(values[i] / matrix.get_sum(keys[i])) * (values[i] / sum_key) for i in range(len(keys))]) sort = np.argsort(-normalized_values) return keys[sort], normalized_values[sort] def forward_normalized_score(key, matrix): sum_key = matrix.get_fw_sum(key) keys, values = matrix.get_forward_adjacent(key) normalized_values = np.array([(values[i] / matrix.get_bw_sum(keys[i])) * (values[i] / sum_key) for i in range(len(keys))]) sort = np.argsort(-normalized_values) return keys[sort], normalized_values[sort] def backward_normalized_score(key, matrix): sum_key = matrix.get_bw_sum(key) keys, values = matrix.get_backward_adjacent(key) normalized_values = np.array([(values[i] / matrix.get_fw_sum(keys[i])) * (values[i] / sum_key) for i in range(len(keys))]) sort = np.argsort(-normalized_values) return keys[sort], normalized_values[sort] # ## Helper class for instructions class RecipeInstructionState(object): def __init__(self): self.current_step = 1 self.id_to_state = {} self.instructions_by_step = {} self.step_by_nodeid = {} self.text_by_nodeid = {} self.ingredients = set() def _add_instruction(self, node_id): s = self.text_by_nodeid[node_id] self.instructions_by_step[self.current_step] = s self.step_by_nodeid[node_id] = self.current_step self.current_step += 1 return self.current_step - 1 def add_text(self, node_id, text, is_instruction=False, is_ingredient=False): self.text_by_nodeid[node_id] = text if is_instruction: self._add_instruction(node_id) if is_ingredient: self.ingredients.add(text) def exists_any_instruction(self, node_ids:list): """check if any instruction exists for list of id's """ for node_id in node_ids: if node_id in self.step_by_nodeid: return True return False def to_markdown(self): md_text = "**Ingredients**:\n" for ing in self.ingredients: md_text += f" * {ing}\n" md_text += "\n\n**Instructions**:\n\n" md_text += "| Step | Instruction |\n" md_text += "| ----:|:----------- |\n" for step in range(1, self.current_step): md_text += f"| {step} | {self.instructions_by_step[step]} |\n" return Markdown(md_text) # ## Recipe Tree # ### Tree Node Base Class class RecipeTreeNode(object): id = 0 def __init__(self, name, constant=False, single_child=False): self._constant = constant self._name = name self._parent = None self._id = str(RecipeTreeNode.id) RecipeTreeNode.id += 1 self._single_child = single_child if self._single_child: self._child = None def child(): return self._child def remove_child(c): assert c == self._child self._child._parent = None self._child = None def childs(): c = self.child() if c is None: return set() return set([c]) def add_child(n): self._child = n n._parent = self self.child = child self.childs = childs self.add_child = add_child self.remove_child = remove_child else: self._childs = set() def childs(): return self._childs def add_child(n): self._childs.add(n) n._parent = self def remove_child(c): assert c in self._childs c._parent = None self._childs.remove(c) self.childs = childs self.add_child = add_child self.remove_child = remove_child def parent(self): return self._parent def root(self): if self._parent is None: return self return self._parent.root() def name(self): return self._name def traverse(self): l = [] for c in self.childs(): l += c.traverse() return [self] + l def traverse_ingredients(self): ingredient_set = [] for c in self.childs(): ingredient_set += c.traverse_ingredients() return ingredient_set def remove(self): p = self.parent() childs = self.childs().copy() assert p is None or not (len(childs) > 1 and p._single_child) for c in childs: self.remove_child(c) if p is not None: p.remove_child(self) if self._single_child and self._child is not None and p._name == self._child._name: # two adjacent nodes with same name would remain after deletion. # merge them! (by adding the child's childs to our parent instead of our childs) childs = self._child.childs() self._child.remove() for c in childs: p.add_child(c) def insert_before(self, n): p = self._parent if p is not None: p.remove_child(self) p.add_child(n) n.add_child(self) def mutate(self): n_node = self.n_node_mutate_options() n_edge = self.n_edge_mutate_options() choice = random.choice(range(n_node + n_edge)) if choice < n_node: self.mutate_node() else: self.mutate_edges() def mutate_edges(self): ings = self.traverse_ingredients() ing = random.choice(ings) a, w = get_backward_adjacent(ing._base_ingredient, m_base_act, c_base_act) if len(a) > 0: action = ea_tools.wheel_of_fortune_selection(a,w) self.insert_before(ActionNode(action)) else: print("Warning: cannot find matching action node for mutation") def mutate_node(self): raise NotImplementedError def n_node_mutate_options(self): return 0 if self._constant else 1 def n_edge_mutate_options(self): n = 1 if self._parent is not None else 0 return n def n_mutate_options(self): return self.n_edge_mutate_options() + self.n_node_mutate_options() def dot_node(self, dot): raise NotImplementedError() def dot(self, d=None): if d is None: d = Digraph() self.dot_node(d) else: self.dot_node(d) if self._parent is not None: d.edge(self._parent._id, self._id) for c in self.childs(): c.dot(d) return d def simplify(self): # simplify nodes (mainly used to delete doubled Mix Nodes) for c in self.childs().copy(): c.simplify() def serialize(self): r = {} r['type'] = str(self.__class__.__name__) r['id'] = self._id r['parent'] = self._parent._id if self._parent is not None else None r['name'] = self._name r['childs'] = [c._id for c in self.childs()] r['constant'] = self._constant r['single_child'] = self._single_child return r def serialize_subtree(self): return [n.serialize() for n in self.traverse()] def node_score(self): raise NotImplementedError() def to_instruction(self, state:RecipeInstructionState): # create an instruction out of a recipe Tree raise NotImplementedError() # ### Mix Node # For the Node Score: just make a simple lookup whether this combination is seen or not. So the node Score is defined as: # class MixNode(RecipeTreeNode): def __init__(self, constant=False): super().__init__("mix", constant, single_child=False) def dot_node(self, dot): dot.node(self._id, label=f"< {self._name}
node score: {self.node_score():.4f}>", shape="diamond", style="filled", color="#d5e8d4") def split(self, set_above, set_below, node_between): assert len(set_above.difference(self.childs())) == 0 assert len(set_below.difference(self.childs())) == 0 n_above = MixNode() n_below = MixNode() p = self.parent() for c in self.childs().copy(): self.remove_child(c) self.remove() for c in set_below: n_below.add_child(c) for c in set_above: n_above.add_child(c) n_above.add_child(node_between) node_between.add_child(n_below) if p is not None: p.add_child(n_above) # test whether the mix nodes are useless if len(n_above.childs()) == 1: n_above.remove() if len(n_below.childs()) == 1: n_below.remove() def n_node_mutate_options(self): return 0 if self._constant or len(self.childs()) <= 2 else len(self.childs()) def mutate_node(self): childs = self.childs() if len(childs) <= 2: print("Warning: cannot modify mix node") return childs = random.sample(childs, len(childs)) n = random.choice(range(1, len(childs)-1)) ings = self.traverse_ingredients() ing = random.choice(ings) base_ing = ing._base_ingredient act = None try: a, w = m_base_act.get_backward_adjacent(base_ing) act = ea_tools.wheel_of_fortune_selection(a,w) except ValueError: print("Warning: cannot mutate given node") if act is not None: between_node = ActionNode(act) self.split(set(childs[:n]), set(childs[n:]), between_node) def node_score(self): child_ingredients = [c.traverse_ingredients() for c in self.childs()] tmp_set = set() cumulative_sets = [] pairwise_tuples = [] for c in child_ingredients: if len(tmp_set) > 0: cumulative_sets.append(tmp_set) pairwise_tuples += [x for x in itertools.product(tmp_set, c)] tmp_set = tmp_set.union(set(c)) s_base = 0 s = 0 for ing_a, ing_b in pairwise_tuples: try: #s_base += sym_score(ing_a._base_ingredient, ing_b._base_ingredient, m_base_mix, c_base_mix) #s += sym_score(ing_a.to_json(), ing_b.to_json(), m_mix, c_mix) # old method: #p1 = sym_p_a_given_b(ing_a.to_json(), ing_b.to_json(), m_mix, c_mix) #p2 = sym_p_a_given_b(ing_b.to_json(), ing_a.to_json(), m_mix, c_mix) #s += 0.5 * p1 + 0.5 * p2 #grouped_ing_a = to_grouped_ingredient(ing_a) #grouped_ing_b = to_grouped_ingredient(ing_b) #ia = m_grouped_mix._label_index[grouped_ing_a.to_json()] #ib = m_grouped_mix._label_index[grouped_ing_b.to_json()] #if c_grouped_mix[ia,ib] > 0 or c_grouped_mix[ib,ia] > 0: # s += 1 ia = m_mix._label_index[ing_a.to_json()] ib = m_mix._label_index[ing_b.to_json()] if c_mix[ia,ib] > 0 or c_mix[ib,ia] > 0: s += 1 except KeyError as e: pass #s_base /= len(pairwise_tuples) s /= len(pairwise_tuples) #return 0.5 * (s_base + s) return s def simplify(self): for c in self.childs().copy(): c.simplify() # if our parent is also a Mix Node, we can just delete ourselve p = self.parent() if p is not None: if type(p) == MixNode: # just delete ourselve self.remove() def to_instruction(self, state:RecipeInstructionState = None): """ returns a RecipeInstructionState """ def english_enum(items, use_and=True): if len(items) > 1 and use_and: return ", ".join(items[:-1]) + " and " + items[-1] return ", ".join(items) if state is None: state = RecipeInstructionState() for c in self.childs(): c.to_instruction(state) text = "" # children with instructions instruction_childs = [] # children without instructions base_childs = [] # childre without instructions that are ingredients ingredient_childs = [] for c in self.childs(): assert type(c) != MixNode if type(c) == IngredientNode: ingredient_childs.append(c) elif c._id not in state.step_by_nodeid: # action node with no step so far, so a base child base_childs.append(c) else: instruction_childs.append(c) if len(base_childs) > 0: use_and= len(ingredient_childs)==0 and len(instruction_childs)==0 text = english_enum([state.text_by_nodeid[c._id] for c in base_childs], use_and=use_and) if len(ingredient_childs) > 0: if len(base_childs) > 0: text += " and mix it with " + english_enum([state.text_by_nodeid[c._id] for c in ingredient_childs]) else: text = "Mix " + english_enum([state.text_by_nodeid[c._id] for c in ingredient_childs]) if len(instruction_childs) > 0: if len(base_childs) == 0 and len(ingredient_childs) == 0: text = "Mix together the results of " else: text += " and mix it together with the results of " text += english_enum([f"step {state.step_by_nodeid[c._id]}" for c in instruction_childs]) text += "." if type(self.parent()) == ActionNode: state.add_text(self._id, text, is_instruction=False) else: state.add_text(self._id, text, is_instruction=True) return state # ### Ingredient Node Class n_wanted_actions = 2 gaussian_normalize_factor = 1 / gaussian(n_wanted_actions, n_wanted_actions, 1) class IngredientNode(RecipeTreeNode): def __init__(self, name, constant=False): super().__init__(name, constant, single_child=True) def get_actions(self): a_list = [] n = self.parent() while n is not None: if type(n) == ActionNode: a_list.append(n.name()) n = n.parent() return a_list def mutate_node(self): if self._constant: return mixes, weights = m_base_mix.get_adjacent(self._name) self._name = ea_tools.wheel_of_fortune_selection(mixes, weights) #self._name = random.choice(base_ingredients) #TODO: change w.r.t. mixing probabilities def traverse_ingredients(self): return [Ingredient(self._name)] def duplicate_actions_score(self, actions): if len(actions) == 0: return 1 seen_actions = set() n_duplicates = 0 for act in actions: if act in seen_actions: n_duplicates += 1 else: seen_actions.add(act) duplicate_actions_score = len(seen_actions) / len(actions) return duplicate_actions_score def duplicate_groups_score(self, actions): if len(actions) == 0: return 1 groups = [AG.groups[a] for a in actions] groups_set = set(groups) return len(groups_set) / len(groups) def node_score(self): actions = self.get_actions() ing = Ingredient(self._name) for a in actions: ing.apply_action(a) heat = heat_score(ing) prepare = prepare_score(ing) score = (heat + prepare) / 2 score *= self.duplicate_actions_score(actions) return score """ actions = self.get_actions() if len(actions) == 0: if p_ingredient_unprepared(self._name) < 0.2: return 0 return 1 seen_actions = set() n_duplicates = 0 for act in actions: if act in seen_actions: n_duplicates += 1 else: seen_actions.add(act) duplicate_actions_score = len(seen_actions) / len(actions) return duplicate_actions_score """ def dot_node(self, dot): dot.node(self._id, label=f"< {self._name}
node score:{self.node_score():.4f}>", shape="box", style="filled", color="#ffe6cc") def to_instruction(self, state:RecipeInstructionState = None): state.add_text(self._id, self._name, is_instruction=False, is_ingredient=True) return state # ### Action Node Class class ActionNode(RecipeTreeNode): def __init__(self, name, constant=False): super().__init__(name, constant, single_child=True) def n_node_mutate_options(self): # beacause we can change or remove ourselve! return 0 if self._constant else 2 def mutate_node(self): if random.choice(range(2)) == 0: # change action ings = self.traverse_ingredients() ing = np.random.choice(ings) base_ing = ing._base_ingredient try: a, w = m_base_act.get_backward_adjacent(base_ing) self._name = ea_tools.wheel_of_fortune_selection(a,w) except ValueError: print("Warning: cannot mutate given node") else: # delete self.remove() def traverse_ingredients(self): ingredient_set = super().traverse_ingredients() for ing in ingredient_set: ing.apply_action(self._name) return ingredient_set def node_score(self): ings = self.child().traverse_ingredients() s = 0 for ing in ings: try: i_act = m_act._x_label_index[self.name()] i_ing = m_act._y_label_index[ing.to_json()] if c_act[i_act,i_ing] > 0: s += 1 except KeyError as e: #print(f"WARNING: no entry found for: {str(e)}") pass ''' # old method: for ing in ings: try: #score = asym_score(self._name, ing.to_json(), m_act, c_act) #base_score = asym_score(self._name, ing._base_ingredient, m_base_act, c_base_act) score = fw_p_a_given_b(self._name, ing._base_ingredient, m_base_act, c_base_act) s += score except KeyError as e: pass ''' return s / len(ings) def dot_node(self, dot): dot.node(self._id, label=f"< {self._name}
node score: {self.node_score():.4f}>", shape="ellipse", style="filled", color="#dae8fc") def to_instruction(self, state:RecipeInstructionState = None): if state is None: state = RecipeInstructionState() for c in self.childs(): c.to_instruction(state) c = self._child if type(c) == MixNode: text = state.text_by_nodeid[c._id] + f" Then {self._name} it." state.add_text(self._id, text, True) elif type(c) == IngredientNode: text = f"{self._name} {state.text_by_nodeid[c._id]}" state.add_text(self._id, text, False) elif type(c) == ActionNode: if c._id in state.step_by_nodeid: text = f"{self._name} the result of step {state.step_by_nodeid[c._id]}" else: prev_words = state.text_by_nodeid[c._id].split() text = f"{prev_words[0]} and {self._name} {' '.join(prev_words[1:])}" state.add_text(self._id, text, True) return state # ### Tree Class class Tree(object): @staticmethod def build_initial_tree(ingredients: list, main_ingredients: list, max_n = 20, wheel_turns = 2): assert set(main_ingredients).issubset(set(ingredients)) def does_action_match(ingredient:str, action:str, t = 0.6): return relative_action_rank(ingredient, action) > t # choose randomly an action for each ingredient by the "wheel of fortune" method actions_for_ing = {} for ing in ingredients: actions_for_ing[ing] = set() action_set, action_weights = m_base_act.get_backward_adjacent(ing) if random_heated(ing): #print(action_set) action_set, action_weights = filter_set_by_group(action_set, action_weights, "heat") #print(action_set) for i in range(wheel_turns): if ing in main_ingredients: # if main ingredient: choose by action probability w = np.array(list(action_weights), dtype=float) w *= (1.0 / np.sum(w)) action = np.random.choice(list(action_set), size=1, replace=False, p=w)[0] else: # else: choose rank based action = ea_tools.wheel_of_fortune_selection(action_set[:max_n], action_weights[:max_n]) actions_for_ing[ing].add(action) #print(f"action {action} for ing {ing}") #print(ing, action) # create ingredient nodes: ingredient_nodes = {} # create ingredient nodes: for ing in ingredients: new_node = IngredientNode(ing, constant=False) # check if we should do a preparation step if random_prepare(ing): # choose a preparation cooking action action_set, action_weights = m_act.get_backward_adjacent(Ingredient(ing).to_json()) action_set, action_weights = filter_set_by_group(action_set, action_weights, "prepare") if len(action_set) > 0: action = ea_tools.wheel_of_fortune_selection(action_set[:max_n], action_weights[:max_n]) act_node = ActionNode(action) act_node.add_child(new_node) new_node = act_node ingredient_nodes[ing] = new_node # starting now with the actions found for the main ingredients and try to match all ingredients together # with that: unprocessed_ings = set(filter(lambda x: len(actions_for_ing[x]) > 0, ingredients)) unprocessed_main_ings = set(filter(lambda x: len(actions_for_ing[x]) > 0, main_ingredients)) while len(unprocessed_main_ings) > 0: main_ing = unprocessed_main_ings.pop() # random action for that ing: act = actions_for_ing[main_ing].pop() act_node = ActionNode(act) mix_node = MixNode() mix_node.add_child(ingredient_nodes[main_ing]) act_node.add_child(mix_node) ingredient_nodes[main_ing] = act_node unprocessed_ings.remove(main_ing) for ing in unprocessed_ings.copy(): if does_action_match(ing, act): mix_node.add_child(ingredient_nodes[ing]) ingredient_nodes[ing] = act_node unprocessed_ings.remove(ing) if ing in unprocessed_main_ings: unprocessed_main_ings.remove(ing) if len(mix_node.childs()) == 1: mix_node.remove() # now make the same with all remaining ingredients: while len(unprocessed_ings) > 0: current_ing = unprocessed_ings.pop() # random action for that ing: act = actions_for_ing[current_ing].pop() act_node = ActionNode(act) mix_node = MixNode() mix_node.add_child(ingredient_nodes[current_ing]) act_node.add_child(mix_node) ingredient_nodes[current_ing] = act_node for ing in unprocessed_ings.copy(): if does_action_match(ing, act): mix_node.add_child(ingredient_nodes[ing]) ingredient_nodes[ing] = act_node unprocessed_ings.remove(ing) if len(mix_node.childs()) == 1: mix_node.remove() root_layer = set([n.root() for n in ingredient_nodes.values()]) root_layer_without_parents = [] for node in root_layer: if node.parent() is None: root_layer_without_parents.append(node) if len(root_layer_without_parents) == 1: root_node = root_layer_without_parents[0] else: root_node = MixNode() for r in root_layer_without_parents: root_node.add_child(r) return root_node @staticmethod def find_ingredients(constant_ingredients, main_ingredients, min_additional:int, max_additional:int, top_ings:int=3, ing_range=50): ''' create an initial set of ingredients, based on given constant ingredients. min_additional and max_additional gives the range of ingredients that are added to our set ''' seen_items = set(constant_ingredients) items = [] scores = [] assert set(main_ingredients).issubset(set(constant_ingredients)) # additional ingredients are choosen w.r.t all given ingredients n_additional_ings = np.random.randint(min_additional, max_additional + 1) # extra ings are ingredients choosen specially for the main ingredient n_extra_ings = int((len(main_ingredients) / len(constant_ingredients)) * n_additional_ings) if n_extra_ings > n_additional_ings: n_extra_ings = n_additional_ings # choose extra ingredients extra_candidates = [] extra_weights = [] for ing in main_ingredients: candidates, weights = normalized_score(ing, m_base_mix) extra_candidates.append(candidates[:ing_range]) extra_weights.append(weights[:ing_range]) extra_ingredients = ea_tools.combined_wheel_of_fortune_selection(extra_candidates, extra_weights, n_extra_ings) for ing in constant_ingredients: # find best matching ingredients best_items = [] best_scores = [] candidates, weights = m_base_mix.get_adjacent(ing) i = 0 while i < len(candidates) and len(best_items) < top_ings: if candidates[i] not in seen_items: best_items.append(candidates[i]) best_scores.append(weights[i]) i += 1 items.append(best_items) scores.append(best_scores) #TODO: error handling if too few options are availabale! additional_ingredients = ea_tools.combined_wheel_of_fortune_selection(items, scores, n_additional_ings - n_extra_ings) return list(constant_ingredients) + list(additional_ingredients) + list(extra_ingredients) @staticmethod def from_ingredients(ingredients: list, main_ingredients: list, min_additional=0, max_additional=10): root = None constant_ingredients = ingredients if max_additional > 0: ingredients = Tree.find_ingredients(ingredients, main_ingredients, min_additional=min_additional, max_additional=max_additional) root = Tree.build_initial_tree(ingredients, main_ingredients) # mark initial ingredient nodes as constant: nodes = root.traverse() for node in nodes: if type(node) == IngredientNode: if node.name() in constant_ingredients: node._constant = True return Tree(root, main_ingredients) @staticmethod def from_serialization(s, main_ingredients = None): def empty_node(raw_n): if raw_n['type'] == "MixNode": node = MixNode(raw_n['constant']) elif raw_n['type'] == "IngredientNode": node = IngredientNode(raw_n['name'], raw_n['constant']) elif raw_n['type'] == "ActionNode": node = ActionNode(raw_n['name'], raw_n['constant']) else: print("unknown node detected") return return node nodes = {} for n in s: nodes[n['id']] = empty_node(n) for n in s: childs = n['childs'] id = n['id'] for c in childs: nodes[id].add_child(nodes[c]) return Tree(nodes[s[0]['id']], main_ingredients) def __init__(self, root, main_ingredients=None): # create a dummy entry node self._root = RecipeTreeNode("root", single_child=True) self._root.add_child(root) self._touched = True self._main_ingredients = main_ingredients def root(self): return self._root.child() def mutate(self): self._touched = True nodes = self.root().traverse() weights = [n.n_mutate_options() for n in nodes] n = random.choices(nodes, weights)[0] n.mutate() # check for simplification after modification self.root().simplify() def dot(self): return self.root().dot() def serialize(self): return [n.serialize() for n in self.root().traverse()] def structure_score(self): n_duplicates = 0 def collect_scores(self): self._mix_scores = [] self._mix_weights = [] self._act_scores = [] self._ing_scores = [] nodes = self.root().traverse() self._n_mix_nodes = 0 self._n_act_nodes = 0 self._n_ing_nodes = 0 s = 0 for n in nodes: if type(n) == MixNode: self._mix_scores.append(n.node_score()) self._mix_weights.append(len(n.childs())) self._n_mix_nodes += 1 if type(n) == ActionNode: self._act_scores.append(n.node_score()) self._n_act_nodes += 1 if type(n) == IngredientNode: self._ing_scores.append(n.node_score()) self._n_ing_nodes += 1 seen_ingredients = set() self._n_duplicates = 0 for n in nodes: if type(n) == IngredientNode: if n.name() in seen_ingredients: self._n_duplicates += 1 else: seen_ingredients.add(n.name()) self._mix_scores = np.array(self._mix_scores) self._mix_weights = np.array(self._mix_weights) self._act_scores = np.array(self._act_scores) self._ing_scores = np.array(self._ing_scores) def mix_scores(self): return self._mix_scores, self._mix_weights def action_scores(self): return self._act_scores def ing_scores(self): return self._ing_scores def main_ingredient_score(self): if self._main_ingredients is None: return 1 ings = self.root().traverse_ingredients() actions_for_ing = {} score_for_ing = {} for ing in ings: if ing._base_ingredient in self._main_ingredients: actions_for_ing[ing._base_ingredient] = ing._action_set score_for_ing[ing._base_ingredient] = 0 for ing in self._main_ingredients: for act in actions_for_ing[ing]: s = fw_p_a_given_b(act, ing, m_base_act, c_base_act) if s > 0.5: score_for_ing[ing] = 1 return sum([score_for_ing[ing] for ing in self._main_ingredients]) / len(self._main_ingredients) def score(self): if not self._touched: return self._score self.collect_scores() s_mix, s_mix_weights = self.mix_scores() s_act = self.action_scores() s_ing = self.ing_scores() #n = len(s_mix) + len(s_act) + len(s_ing) avg_mix = np.average(s_mix) if len(s_mix) > 0 else 1 avg_act = np.average(s_act) if len(s_act) > 0 else 1 avg_ing = np.average(s_ing) if len(s_ing) > 0 else 1 sum_mix = np.sum(s_mix) if len(s_mix) > 0 else 0 sum_act = np.sum(s_act) if len(s_act) > 0 else 0 sum_ing = np.sum(s_ing) if len(s_ing) > 0 else 0 self._touched = False contains_main_ingred = True base_main_ings = [i._base_ingredient for i in self.root().traverse_ingredients()] for ing in self._main_ingredients: if ing not in base_main_ings: contains_main_ingred = False self._score = 0 break if contains_main_ingred: # boost creativity if len(s_act) < 3: self._score = 0 elif len(s_ing) < 3: self._score = 0 else: weighted_mix_score = np.array([s_mix[i] * s_mix_weights[i] for i in range(len(s_mix))]) #print(weighted_mix_score) n = len(s_act) + len(s_ing) + np.sum(s_mix_weights) self._score = (np.sum(weighted_mix_score) + sum_act + sum_ing) / n #print(self._score) self._score *= (len(s_ing) - self._n_duplicates) / len(s_ing) #self._score = 0.95 * self._score + 0.05 * self.main_ingredient_score() return self._score def copy(self): return Tree.from_serialization(self.serialize(), self._main_ingredients) # ## Population class Population(object): def __init__(self, start_ingredients, main_ingredients, n_population = 50, min_additional=0, max_additional=15, mutations=3): self.population = [] for i in tqdm(range(n_population), desc="build initial population"): self.population.append(Tree.from_ingredients(start_ingredients, main_ingredients, min_additional=min_additional, max_additional=max_additional)) self._n = n_population self._n_mutations = mutations def mutate(self): for tree in self.population.copy(): t_clone = tree.copy() for i in range(self._n_mutations): t_clone.mutate() #t_clone.mutate() #t_clone.mutate() self.population.append(t_clone) def pairwise_competition(self): new_population = [] indices = list(range(len(self.population))) random.shuffle(indices) for i in range(len(self.population) // 2): i_a = indices[2*i] i_b = indices[2*i+1] if self.population[i_a].score() > self.population[i_b].score(): new_population.append(self.population[i_a]) else: new_population.append(self.population[i_b]) self.population = new_population def crossover(self): # shuffle indices indices = list(range(len(self.population) // 2)) indices = [i + len(self.population) // 2 for i in indices] random.shuffle(indices) # perform crossover for random pairs for i in range(len(self.population) // 4): i_a = indices[2*i] i_b = indices[2*i+1] self.pairwise_crossover(self.population[i_a], self.population[i_b]) def pairwise_crossover(self, tree_a, tree_b): # for crossover: find a random subtree in both trees, and switch them # first, select one random mix node from both a_nodes = tree_a.root().traverse() b_nodes = tree_b.root().traverse() a_mix_nodes = [] b_mix_nodes = [] for n in a_nodes: if type(n) == MixNode: a_mix_nodes.append(n) for n in b_nodes: if type(n) == MixNode: b_mix_nodes.append(n) a_mix_node = np.random.choice(a_mix_nodes) b_mix_node = np.random.choice(b_mix_nodes) # now select one random child, we will switch the subtrees there a_child = np.random.choice(list(a_mix_node.childs())) b_child = np.random.choice(list(b_mix_node.childs())) # ...and perform the switch # manually remove references a_mix_node.remove_child(a_child) b_mix_node.remove_child(b_child) # and add child to other subtree a_mix_node.add_child(b_child) b_mix_node.add_child(a_child) def hold_best(self, n=10): scores = np.array([tree.score() for tree in self.population]) sorted_indices = np.argsort(-scores) self.population = np.array(self.population)[sorted_indices[:n]].tolist() def run(self, n=50): avg_scores = [] for i in tqdm(range(n), desc="run evolutionary cycles"): self.mutate() self.crossover() self.pairwise_competition() #self.collect_scores() #self.hold_best(self._n) scores = [t.score() for t in self.population] avg_scores.append(scores) return avg_scores def plot_population(self, n_best=10): scores = [tree.score() for tree in self.population] ii = np.argsort(-np.array(scores))[:n_best] for i in ii: self.population[i].root().simplify() display(self.population[i].root().dot()) display(Markdown(f"**Recipe Score**: {scores[i]}")) display(self.population[i].root().to_instruction().to_markdown()) # ## Run Evolutionary Algorithm p = Population(["noodle"],['noodle'], min_additional=4, max_additional=13, n_population = 50, mutations=1) #p_ingredient_unprepared(list(p.population[0].root().childs())[0]._name) < 0.2 #avg = p.run(10) #p.plot_population(n_best=20)