又搞到3点，还是没有找到问题在哪，不过已经排除了是forward的问题

algorithms/neat/genome/numpy/__init__.py

@@ -1,4 +1,4 @@
-from .genome import create_initialize_function, expand, expand_single, analysis
+from .genome import create_initialize_function, expand, expand_single, analysis, pop_analysis
 from .distance import distance
 from .mutate import create_mutate_function
 from .forward import create_forward_function

algorithms/neat/genome/numpy/aggregations.py

@@ -69,7 +69,6 @@ agg_name2key = {
 
 def agg(idx, z):
     idx = np.asarray(idx, dtype=np.int32)
-
     if np.all(z == 0.):
         return 0
     else:

algorithms/neat/genome/numpy/forward.py

@@ -76,7 +76,6 @@ def forward_single(inputs: NDArray, N: int, input_idx: NDArray, output_idx: NDAr
         # for some nodes (inputs nodes), the output z will be nan, thus we do not update the vals
         ini_vals[i] = z
 
-
     return ini_vals[output_idx]
 
 

algorithms/neat/genome/numpy/genome.py

@@ -198,8 +198,13 @@ def analysis(nodes: NDArray, connections: NDArray, input_keys, output_keys) -> \
 
 def pop_analysis(pop_nodes, pop_connections, input_keys, output_keys):
     res = []
+    total_nodes, total_connections = 0, 0
     for nodes, connections in zip(pop_nodes, pop_connections):
-        res.append(analysis(nodes, connections, input_keys, output_keys))
+        nodes, connections = analysis(nodes, connections, input_keys, output_keys)
+        res.append((nodes, connections))
+        total_nodes += len(nodes)
+        total_connections += len(connections)
+    print(total_nodes - 200, total_connections)
     return res
 
 

algorithms/neat/genome/numpy/mutate.py

@@ -9,6 +9,8 @@ from .utils import fetch_random, fetch_first, I_INT
 from .genome import add_node, add_connection_by_idx, delete_node_by_idx, delete_connection_by_idx
 from .graph import check_cycles
 
+add_node_cnt, delete_node_cnt, add_connection_cnt, delete_connection_cnt = 0, 0, 0, 0
+
 
 def create_mutate_function(config, input_keys, output_keys, batch: bool):
     """
@@ -79,11 +81,15 @@ def create_mutate_function(config, input_keys, output_keys, batch: bool):
         return mutate_func
     else:
         def batch_mutate_func(pop_nodes, pop_connections, new_node_keys):
+            global add_node_cnt, delete_node_cnt, add_connection_cnt, delete_connection_cnt
+            add_node_cnt, delete_node_cnt, add_connection_cnt, delete_connection_cnt = 0, 0, 0, 0
             res_nodes, res_connections = [], []
             for nodes, connections, new_node_key in zip(pop_nodes, pop_connections, new_node_keys):
                 nodes, connections = mutate_func(nodes, connections, new_node_key)
                 res_nodes.append(nodes)
                 res_connections.append(connections)
+            # print(f"add_node_cnt: {add_node_cnt}, delete_node_cnt: {delete_node_cnt}, "
+            #       f"add_connection_cnt: {add_connection_cnt}, delete_connection_cnt: {delete_connection_cnt}")
             return np.stack(res_nodes, axis=0), np.stack(res_connections, axis=0)
 
         return batch_mutate_func
@@ -161,6 +167,8 @@ def mutate(nodes: NDArray,
     :return:
     """
 
+    global add_node_cnt, delete_node_cnt, add_connection_cnt, delete_connection_cnt
+
     # mutate_structure
     def nothing(n, c):
         return n, c
@@ -200,18 +208,22 @@ def mutate(nodes: NDArray,
         # mutate add node
         if rand() < add_node_rate:
             nodes, connections = m_add_node(nodes, connections)
+            add_node_cnt += 1
 
         # mutate delete node
         if rand() < delete_node_rate:
             nodes, connections = m_delete_node(nodes, connections)
+            delete_node_cnt += 1
 
         # mutate add connection
         if rand() < add_connection_rate:
             nodes, connections = m_add_connection(nodes, connections)
+            add_connection_cnt += 1
 
         # mutate delete connection
         if rand() < delete_connection_rate:
             nodes, connections = m_delete_connection(nodes, connections)
+            delete_connection_cnt += 1
 
     nodes, connections = mutate_values(nodes, connections, bias_mean, bias_std, bias_mutate_strength,
                                        bias_mutate_rate, bias_replace_rate, response_mean, response_std,
@@ -220,6 +232,8 @@ def mutate(nodes: NDArray,
                                        weight_mutate_rate, weight_replace_rate, act_range, act_replace_rate, agg_range,
                                        agg_replace_rate, enabled_reverse_rate)
 
+    # print(add_node_cnt, delete_node_cnt, add_connection_cnt, delete_connection_cnt)
+
     return nodes, connections
 
 
@@ -321,9 +335,9 @@ def mutate_float_values(old_vals: NDArray, mean: float, std: float,
     replace = np.random.normal(size=old_vals.shape) * std + mean
     r = rand(*old_vals.shape)
     new_vals = old_vals
-    new_vals = np.where(r < mutate_rate, new_vals + noise, new_vals)
+    new_vals = np.where(r <= mutate_rate, new_vals + noise, new_vals)
     new_vals = np.where(
-        np.logical_and(mutate_rate < r, r < mutate_rate + replace_rate),
+        (mutate_rate < r) & (r <= mutate_rate + replace_rate),
         replace,
         new_vals
     )
@@ -413,7 +427,7 @@ def mutate_delete_node(nodes: NDArray, connections: NDArray,
     node_key, node_idx = choice_node_key(nodes, input_keys, output_keys,
                                          allow_input_keys=False, allow_output_keys=False)
 
-    if np.isnan(node_key):
+    if node_idx == I_INT:
         return nodes, connections
 
     # delete the node

algorithms/neat/genome/origin_neat/__init__.py

@@ -0,0 +1,82 @@
+import numpy as np
+from .genome import Genome
+from .gene import NodeGene, ConnectionGene
+from .feedforward import FeedForwardNetwork
+
+def object2array(genome, N):
+    """
+    convert objective genome to array
+    :param genome:
+    :param N: the size of the array
+    :return: Tuple(Array, Array), represents the nodes and connections array
+    nodes: shape(N, 5), dtype=float
+    connections: shape(2, N, N), dtype=float
+    con[:, i, j] != nan, means there is a connection from i to j
+    """
+    nodes = np.full((N, 5), np.nan)
+    connections = np.full((2, N, N), np.nan)
+
+    assert len(genome.nodes) + len(genome.input_keys) + 1 <= N  # remain one inf row for mutation adding extra node
+
+    idx = 0
+    n2i = {}
+    for i in genome.input_keys:
+        nodes[idx, 0] = i
+        n2i[i] = idx
+        idx += 1
+
+    for k, v in genome.nodes.items():
+        nodes[idx, 0] = k
+        nodes[idx, 1] = v.bias
+        nodes[idx, 2] = v.response
+        nodes[idx, 3] = 0
+        nodes[idx, 4] = 0
+        n2i[k] = idx
+        idx += 1
+
+    for (f, t), v in genome.connections.items():
+        f_i, t_i = n2i[f], n2i[t]
+        connections[0, f_i, t_i] = v.weight
+        connections[1, f_i, t_i] = v.enabled
+
+    return nodes, connections
+
+
+def array2object(config, nodes, connections):
+    """
+    convert array to genome
+    :param config:
+    :param nodes:
+    :param connections:
+    :return:
+    """
+    genome = Genome(0, config, None, init_val=False)
+    genome.input_keys = [0, 1]
+    genome.output_keys = [2]
+    idx2key = {}
+    for i in range(nodes.shape[0]):
+        key = nodes[i, 0]
+        if np.isnan(key):
+            continue
+        key = int(key)
+        idx2key[i] = key
+        if key in genome.input_keys:
+            continue
+        node_gene = NodeGene(key, config, init_val=False)
+        node_gene.bias = nodes[i, 1]
+        node_gene.response = nodes[i, 2]
+        node_gene.act = 'sigmoid'
+        node_gene.agg = 'sum'
+        genome.nodes[key] = node_gene
+
+    for i in range(connections.shape[1]):
+        for j in range(connections.shape[2]):
+            if np.isnan(connections[0, i, j]):
+                continue
+            key = (idx2key[i], idx2key[j])
+            connection_gene = ConnectionGene(key, config, init_val=False)
+            connection_gene.weight = connections[0, i, j]
+            connection_gene.enabled = connections[1, i, j] == 1
+            genome.connections[key] = connection_gene
+
+    return genome

algorithms/neat/genome/origin_neat/activations.py

@@ -0,0 +1,17 @@
+"""
+Has the built-in activation functions,
+code for using them,
+and code for adding new user-defined ones
+"""
+import math
+
+def sigmoid_activation(z):
+    z = max(-60.0, min(60.0, 5.0 * z))
+    return 1.0 / (1.0 + math.exp(-z))
+
+
+activation_dict = {
+    "sigmoid": sigmoid_activation,
+}
+
+full_activation_list = list(activation_dict.keys())

algorithms/neat/genome/origin_neat/aggregations.py

@@ -0,0 +1,14 @@
+"""
+Has the built-in aggregation functions, code for using them,
+and code for adding new user-defined ones.
+"""
+
+def sum_aggregation(x):
+    return sum(x)
+
+
+aggregation_dict = {
+    'sum': sum_aggregation,
+}
+
+full_aggregation_list = list(aggregation_dict.keys())

algorithms/neat/genome/origin_neat/feedforward.py

@@ -0,0 +1,54 @@
+from .graphs import node_calculate_sequence
+from .activations import activation_dict
+from .aggregations import aggregation_dict
+
+
+class FeedForwardNetwork(object):
+    def __init__(self, inputs, outputs, node_evals):
+        self.input_nodes = inputs
+        self.output_nodes = outputs
+        self.node_evals = node_evals
+        self.values = dict((key, 0.0) for key in inputs + outputs)
+
+    def activate(self, inputs):
+        if len(self.input_nodes) != len(inputs):
+            raise RuntimeError("Expected {0:n} inputs, got {1:n}".format(len(self.input_nodes), len(inputs)))
+
+        for k, v in zip(self.input_nodes, inputs):
+            self.values[k] = v
+
+        for node, act_func, agg_func, bias, response, links in self.node_evals:
+            node_inputs = []
+            for i, w in links:
+                node_inputs.append(self.values[i] * w)
+            if len(node_inputs) == 0:
+                s = 0.0
+            else:
+                s = agg_func(node_inputs)
+            self.values[node] = act_func(bias + response * s)
+
+        return [self.values[i] for i in self.output_nodes]
+
+    @staticmethod
+    def create(genome):
+        """ Receives a genome and returns its phenotype (a FeedForwardNetwork). """
+
+        # Gather expressed connections.
+        connections = [cg.key for cg in genome.connections.values() if cg.enabled]
+
+        seqs, useful_connections = node_calculate_sequence(genome.input_keys, genome.output_keys, connections)
+        node_evals = []
+        for node in seqs:
+            inputs = []
+            for conn_key in useful_connections:
+                inode, onode = conn_key
+                if onode == node:
+                    cg = genome.connections[conn_key]
+                    inputs.append((inode, cg.weight))
+
+            ng = genome.nodes[node]
+            act_func = activation_dict[ng.act]
+            agg_func = aggregation_dict[ng.agg]
+            node_evals.append((node, act_func, agg_func, ng.bias, ng.response, inputs))
+
+        return FeedForwardNetwork(genome.input_keys, genome.output_keys, node_evals)

algorithms/neat/genome/origin_neat/gene.py

@@ -0,0 +1,152 @@
+from typing import Tuple
+from random import gauss, choice, random
+
+
+def clip(x, min_val, max_val):
+    return min(max(x, min_val), max_val)
+
+
+class NodeGene:
+
+    def __init__(self, key: int, config, init_val=True):
+        self.key = key
+        self.config = config
+
+        if init_val:
+            self.init_value()
+        else:
+            self.bias = 0
+            self.response = 1
+            self.act = 0
+            self.agg = 0
+
+    def __repr__(self):
+        return f'node({self.key}, bias: {self.bias:.3f}, ' \
+               f'response: {self.response:.3f}, act: {self.act}, agg: {self.agg})'
+
+    def __eq__(self, other):
+        if not isinstance(other, NodeGene):
+            return False
+        return self.key == other.key and \
+            self.bias == other.bias and \
+            self.response == other.response and \
+            self.act == other.act and \
+            self.agg == other.agg
+
+    def copy(self):
+        new_gene = self.__class__(self.key, config=self.config, init_val=False)
+        new_gene.bias = self.bias  # numpy array is mutable, so we need to copy it
+        new_gene.response = self.response
+        new_gene.act = self.act
+        new_gene.agg = self.agg
+        return new_gene
+
+    def init_value(self):
+        c = self.config.gene
+        self.bias = gauss(c.bias.init_mean, c.bias.init_stdev)
+        self.response = gauss(c.response.init_mean, c.response.init_stdev)
+        self.act = choice(c.activation.options)
+        self.agg = choice(c.aggregation.options)
+
+        self.bias = clip(self.bias, c.bias.min_value, c.bias.max_value)
+        self.response = clip(self.response, c.response.min_value, c.response.max_value)
+
+    def distance(self, other):
+        s = abs(self.bias - other.bias) + abs(self.response - other.response)
+        if self.act != other.act:
+            s += 1
+        if self.agg != other.agg:
+            s += 1
+        return s
+
+    def mutate(self):
+        self.bias = mutate_float(self.bias, self.config.gene.bias)
+        self.response = mutate_float(self.response, self.config.gene.response)
+        self.act = mutate_string(self.act, self.config.gene.activation)
+        self.agg = mutate_string(self.agg, self.config.gene.aggregation)
+
+    @classmethod
+    def crossover(cls, g1, g2):
+        assert g1.key == g2.key
+        c = cls(g1.key, g1.config, init_val=False)
+        c.bias = g1.bias if random() > 0.5 else g2.bias
+        c.response = g1.response if random() > 0.5 else g2.response
+        c.act = g1.act if random() > 0.5 else g2.act
+        c.agg = g1.agg if random() > 0.5 else g2.agg
+        return c
+
+
+class ConnectionGene:
+    def __init__(self, key: Tuple[int, int], config, init_val=True):
+        self.key = key
+        self.config = config
+        self.enabled = True
+        if init_val:
+            self.init_value()
+        else:
+            self.weight = 1
+
+    def __repr__(self):
+        return f'connection({self.key}, {self.weight:.3f}, {self.enabled})'
+
+    def __eq__(self, other):
+        if not isinstance(other, ConnectionGene):
+            return False
+        return self.key == other.key and \
+            self.weight == other.weight and \
+            self.enabled == other.enabled
+
+    def copy(self):
+        new_gene = self.__class__(self.key, self.config, init_val=False)
+        new_gene.weight = self.weight
+        new_gene.enabled = self.enabled
+        return new_gene
+
+    def init_value(self):
+        c = self.config.gene
+        self.weight = gauss(c.weight.init_mean, c.weight.init_stdev)
+        self.weight = clip(self.weight, c.weight.min_value, c.weight.max_value)
+
+    def distance(self, other):
+        s = abs(self.weight - other.weight)
+        if self.enabled != other.enabled:
+            s += 1
+        return s
+
+    def mutate(self):
+        self.weight = mutate_float(self.weight, self.config.gene.weight)
+        if random() < self.config.gene.enabled.mutate_rate:
+            self.enabled = not self.enabled
+
+    @classmethod
+    def crossover(cls, g1, g2):
+        assert g1.key == g2.key
+        c = cls(g1.key, g1.config, init_val=False)
+        c.weight = g1.weight if random() > 0.5 else g2.weight
+        c.enabled = g1.enabled if random() > 0.5 else g2.enabled
+        return c
+
+
+# HAHA, exactly the bug is here!!
+# After I fixed it, the result is much better!!
+def mutate_float(v, vc):
+    """vc -> value config"""
+    r = random()
+    if r < vc.mutate_rate:
+        v += gauss(0, vc.mutate_power)
+        v = clip(v, vc.min_value, vc.max_value)
+    # Previous, seems like a huge bug
+    # if r < vc.mutate_rate + vc.replace_rate:
+    # Now:
+    elif r < vc.mutate_rate + vc.replace_rate:
+        v = gauss(vc.init_mean, vc.init_stdev)
+        v = clip(v, vc.min_value, vc.max_value)
+    return v
+
+
+def mutate_string(v, vc):
+    """vc -> value config"""
+    r = random()
+    if r < vc.mutate_rate:
+        v = choice(vc.options)
+    return v

algorithms/neat/genome/origin_neat/genome.py

@@ -0,0 +1,246 @@
+from random import random, choice
+
+from .gene import NodeGene, ConnectionGene
+from .graphs import creates_cycle
+
+
+class Genome:
+    def __init__(self, key, config, global_idx, init_val=True):
+        # Unique identifier for a genome instance.
+        self.key = key
+        self.config = config
+        self.global_idx = global_idx
+
+        # (gene_key, gene) pairs for gene sets.
+        self.connections = {}
+        self.nodes = {}
+
+        # Fitness results.
+        self.fitness = None
+
+        self.input_keys = [-i - 1 for i in range(config.basic.num_inputs)]
+        self.output_keys = [i for i in range(config.basic.num_outputs)]
+
+        if init_val:
+            self.initialize()
+
+    def __repr__(self):
+        nodes_info = ',\n\t\t'.join(map(str, self.nodes.values()))
+        connections_info = ',\n\t\t'.join(map(str, self.connections.values()))
+
+        return f'Genome(\n\t' \
+               f'key: {self.key}, \n' \
+               f'\tinput_keys: {self.input_keys}, \n' \
+               f'\toutput_keys: {self.output_keys}, \n' \
+               f'\tnodes: \n\t\t' \
+               f'{nodes_info} \n' \
+               f'\tconnections: \n\t\t' \
+               f'{connections_info} \n)'
+
+    def __eq__(self, other):
+        if not isinstance(other, Genome):
+            return False
+        if self.key != other.key:
+            return False
+        if len(self.nodes) != len(other.nodes) or len(self.connections) != len(other.connections):
+            return False
+        for k, v in self.nodes.items():
+            o_v = other.nodes.get(k)
+            if o_v is None or v != o_v:
+                return False
+        for k, v in self.connections.items():
+            o_v = other.connections.get(k)
+            if o_v is None or v != o_v:
+                return False
+        return True
+
+    def initialize(self):
+        """Configure a new genome based on the given configuration."""
+
+        # Create node genes for the output pins.
+        for node_key in self.output_keys:
+            self.nodes[node_key] = NodeGene(node_key, self.config, init_val=True)
+
+        # Add connections based on initial connectivity type.
+        # ONLY ALLOW FULL HERE AND NO HIDDEN!!!
+        for i in self.input_keys:
+            for j in self.output_keys:
+                key = (i, j)
+                self.connections[key] = ConnectionGene(key, self.config, init_val=True)
+
+    def distance(self, other):
+        """Calculate the distance between two genomes."""
+
+        wc = self.config.genome.compatibility_weight_coefficient
+        dc = self.config.genome.compatibility_disjoint_coefficient
+
+        node_distance = 0.0
+        if self.nodes or other.nodes:  # otherwise, both are empty
+            disjoint_nodes = 0
+            for k2 in other.nodes:
+                if k2 not in self.nodes:
+                    disjoint_nodes += 1
+
+            for k1, n1 in self.nodes.items():
+                n2 = other.nodes.get(k1)
+                if n2 is None:
+                    disjoint_nodes += 1
+                else:
+                    # Homologous genes compute their own distance value.
+                    node_distance += n1.distance(n2)
+
+            max_nodes = max(len(self.nodes), len(other.nodes))
+            node_distance = (wc * node_distance + dc * disjoint_nodes) / max_nodes
+
+        connection_distance = 0.0
+        if self.connections or other.connections:
+            disjoint_connections = 0
+            for k2 in other.connections:
+                if k2 not in self.connections:
+                    disjoint_connections += 1
+
+            for k1, c1 in self.connections.items():
+                c2 = other.connections.get(k1)
+                if c2 is None:
+                    disjoint_connections += 1
+                else:
+                    # Homologous genes compute their own distance value.
+                    connection_distance += c1.distance(c2)
+
+            max_conn = max(len(self.connections), len(other.connections))
+            connection_distance = (wc * connection_distance + dc * disjoint_connections) / max_conn
+
+        return node_distance + connection_distance
+
+    @classmethod
+    def crossover(cls, new_key, g1, g2):
+        if g1.fitness > g2.fitness:
+            p1, p2 = g1, g2
+        else:
+            p1, p2 = g2, g1
+
+        child = cls(new_key, p1.config, p1.global_idx, init_val=False)
+
+        for k, cg1 in p1.connections.items():
+            cg2 = p2.connections.get(k)
+            if cg2 is None:
+                child.connections[k] = cg1.copy()
+            else:
+                child.connections[k] = ConnectionGene.crossover(cg1, cg2)
+
+        for k, ng1 in p1.nodes.items():
+            ng2 = p2.nodes.get(k)
+            if ng2 is None:
+                child.nodes[k] = ng1.copy()
+            else:
+                child.nodes[k] = NodeGene.crossover(ng1, ng2)
+
+        return child
+
+    def mutate(self):
+        c = self.config.genome
+
+        if c.single_structural_mutation:
+            div = max(1, c.conn_add_prob + c.conn_delete_prob + c.node_add_prob + c.node_delete_prob)
+            r = random()
+
+            if r < c.node_add_prob / div:
+                self.mutate_add_node()
+            elif r < (c.node_add_prob + c.node_delete_prob) / div:
+                self.mutate_delete_node()
+            elif r < (c.node_add_prob + c.node_delete_prob + c.conn_add_prob) / div:
+                self.mutate_add_connection()
+            elif r < (c.node_add_prob + c.node_delete_prob + c.conn_add_prob + c.conn_delete_prob) / div:
+                self.mutate_delete_connection()
+        else:
+            if random() < c.node_add_prob:
+                self.mutate_add_node()
+            if random() < c.node_delete_prob:
+                self.mutate_delete_node()
+            if random() < c.conn_add_prob:
+                self.mutate_add_connection()
+            if random() < c.conn_delete_prob:
+                self.mutate_delete_connection()
+
+        for cg in self.connections.values():
+            cg.mutate()
+
+        for ng in self.nodes.values():
+            ng.mutate()
+
+    def mutate_add_node(self):
+        # create a node from splitting a connection
+        if not self.connections:
+            return -1
+
+        # Choose a random connection to split
+        conn_to_split = choice(list(self.connections.values()))
+        new_node_id = self.global_idx.next_node()
+        ng = NodeGene(new_node_id, self.config, init_val=False)
+        self.nodes[new_node_id] = ng
+
+        # Create two new connections
+        conn_to_split.enabled = False
+        i, o = conn_to_split.key
+        con1 = ConnectionGene((i, new_node_id), self.config, init_val=False)
+        con2 = ConnectionGene((new_node_id, o), self.config, init_val=False)
+
+        # The new node+connections have roughly the same behavior as
+        # the original connection (depending on the activation function of the new node).
+        con2.weight = conn_to_split.weight
+        self.connections[con1.key] = con1
+        self.connections[con2.key] = con2
+
+        return 1
+
+    def mutate_delete_node(self):
+        # Do nothing if there are no non-output nodes.
+        available_nodes = [k for k in self.nodes if k not in self.output_keys]
+        if not available_nodes:
+            return -1
+
+        del_key = choice(available_nodes)
+        connections_to_delete = set()
+        for k, v in self.connections.items():
+            if del_key in v.key:
+                connections_to_delete.add(v.key)
+
+        for key in connections_to_delete:
+            del self.connections[key]
+
+        del self.nodes[del_key]
+
+        return del_key
+
+    def mutate_add_connection(self):
+        """
+        Attempt to add a new connection, the only restriction being that the output
+        node cannot be one of the network input pins.
+        """
+        possible_outputs = list(self.nodes)
+        out_node = choice(possible_outputs)
+
+        possible_inputs = possible_outputs + self.input_keys
+        in_node = choice(possible_inputs)
+
+        # in recurrent networks, the input node can be the same as the output node
+        key = (in_node, out_node)
+        if key in self.connections:
+            self.connections[key].enabled = True
+            return -1
+
+        # if feedforward network, check if the connection creates a cycle
+        if self.config.genome.feedforward and creates_cycle(self.connections.keys(), key):
+            return -1
+
+        cg = ConnectionGene(key, self.config, init_val=True)
+        self.connections[key] = cg
+        return key
+
+    def mutate_delete_connection(self):
+        if self.connections:
+            key = choice(list(self.connections.keys()))
+            del self.connections[key]
+
+    def complexity(self):
+        return len(self.connections) * 2 + len(self.nodes) * 4

algorithms/neat/genome/origin_neat/graphs.py

@@ -0,0 +1,130 @@
+"""Directed graph algorithm implementations."""
+
+
+def creates_cycle(connections, test):
+    """
+    Returns true if the addition of the 'test' connection would create a cycle,
+    assuming that no cycle already exists in the graph represented by 'connections'.
+    """
+    i, o = test
+    if i == o:
+        return True
+
+    visited = {o}
+    while True:
+        num_added = 0
+        for a, b in connections:
+            if a in visited and b not in visited:
+                if b == i:
+                    return True
+
+                visited.add(b)
+                num_added += 1
+
+        if num_added == 0:
+            return False
+
+
+def required_for_output(inputs, outputs, connections):
+    """
+    Collect the nodes whose state is required to compute the final network output(s).
+    :param inputs: list of the input identifiers
+    :param outputs: list of the output node identifiers
+    :param connections: list of (input, output) connections in the network.
+    NOTE: It is assumed that the input identifier set and the node identifier set are disjoint.
+    By convention, the output node ids are always the same as the output index.
+
+    Returns a set of identifiers of required nodes.
+    """
+    assert not set(inputs).intersection(outputs)
+
+    required = set(outputs)
+    s = set(outputs)
+    while 1:
+        # Find nodes not in s whose output is consumed by a node in s.
+        t = set(a for (a, b) in connections if b in s and a not in s)
+
+        if not t:
+            break
+
+        layer_nodes = set(x for x in t if x not in inputs)
+        if not layer_nodes:
+            break
+
+        required = required.union(layer_nodes)
+        s = s.union(t)
+
+    return required
+
+
+def feed_forward_layers(inputs, outputs, connections):
+    """
+    Collect the layers whose members can be evaluated in parallel in a feed-forward network.
+    :param inputs: list of the network input nodes
+    :param outputs: list of the output node identifiers
+    :param connections: list of (input, output) connections in the network.
+
+    Returns a list of layers, with each layer consisting of a set of node identifiers.
+    Note that the returned layers do not contain nodes whose output is ultimately
+    never used to compute the final network output.
+    """
+
+    required = required_for_output(inputs, outputs, connections)
+
+    layers = []
+    s = set(inputs)
+    while 1:
+        # Find candidate nodes c for the next layer.  These nodes should connect
+        # a node in s to a node not in s.
+        c = set(b for (a, b) in connections if a in s and b not in s)
+        # Keep only the used nodes whose entire input set is contained in s.
+        t = set()
+        for n in c:
+            if n in required and all(a in s for (a, b) in connections if b == n):
+                t.add(n)
+
+        if not t:
+            break
+
+        layers.append(t)
+        s = s.union(t)
+
+    return layers
+
+
+def node_calculate_sequence(inputs, outputs, connections):
+    """
+    Collect the sequence of nodes to calculate in order to compute the final network output(s).
+    :param required_nodes:
+    :param connections:
+    :return:
+    """
+    required_nodes = required_for_output(inputs, outputs, connections)
+    useful_nodes = required_nodes.copy()
+    useful_nodes.update(inputs)
+    useful_connections = [c for c in connections if c[0] in useful_nodes and c[1] in useful_nodes]
+
+    # do topological sort on useful_connections
+    in_degrees = {n: 0 for n in useful_nodes}
+    for a, b in useful_connections:
+        in_degrees[b] += 1
+    topological_order = []
+    while len(topological_order) < len(useful_nodes):
+        for n in in_degrees:
+            if in_degrees[n] == 0:
+                topological_order.append(n)
+                in_degrees[n] -= 1
+                for a, b in useful_connections:
+                    if a == n:
+                        in_degrees[b] -= 1
+
+    [topological_order.remove(n) for n in inputs]  # remove inputs from topological order
+    return topological_order, useful_connections
+
+
+if __name__ == '__main__':
+    inputs = [-1, -2]
+    outputs = [0]
+    connections = [(-2, 2), (-2, 3), (4, 0), (3, 0), (2, 0), (2, 3), (2, 4)]
+    seqs = node_calculate_sequence(inputs, outputs, connections)
+    print(seqs)

algorithms/neat/pipeline.py

@@ -6,7 +6,12 @@ import numpy as np
 from .species import SpeciesController
 from .genome.numpy import create_initialize_function, create_mutate_function, create_forward_function
 from .genome.numpy import batch_crossover
-from .genome.numpy import expand, expand_single
+from .genome.numpy import expand, expand_single, pop_analysis
+
+from .genome.origin_neat import *
+
+xor_inputs = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
+xor_outputs = np.array([[0], [1], [1], [0]])
 
 
 class Pipeline:
@@ -14,8 +19,7 @@ class Pipeline:
     Neat algorithm pipeline.
     """
 
-    def __init__(self, config, seed=42):
-        np.random.seed(seed)
+    def __init__(self, config):
 
         self.config = config
         self.N = config.basic.init_maximum_nodes
@@ -48,6 +52,15 @@ class Pipeline:
         return func
 
     def tell(self, fitnesses):
+        # idx = np.argmax(fitnesses)
+        # print(f"argmax: {idx}, max: {np.max(fitnesses)}, a_max: {fitnesses[idx]}")
+        # n, c = self.pop_nodes[idx], self.pop_connections[idx]
+        # func = create_forward_function(n, c, self.N, self.input_idx, self.output_idx, batch=True)
+        # out = func(xor_inputs)
+        # print(f"max fitness: {fitnesses[idx]}")
+        # print(f"real fitness: {4 - np.sum(np.abs(out - xor_outputs), axis=0)}")
+        # print(f"Out:\n{func(np.array([[0, 0], [0, 1], [1, 0], [1, 1]]))}")
+
         self.generation += 1
 
         self.species_controller.update_species_fitnesses(fitnesses)
@@ -56,12 +69,31 @@ class Pipeline:
 
         self.update_next_generation(crossover_pair)
 
-        # print(pop_analysis(self.pop_nodes, self.pop_connections, self.input_idx, self.output_idx))
+        analysis = pop_analysis(self.pop_nodes, self.pop_connections, self.input_idx, self.output_idx)
+
+        try:
+            for nodes, connections in zip(self.pop_nodes, self.pop_connections):
+                g = array2object(self.config, nodes, connections)
+                print(g)
+                net = FeedForwardNetwork.create(g)
+                real_out = [net.activate(x) for x in xor_inputs]
+                func = create_forward_function(nodes, connections, self.N, self.input_idx, self.output_idx, batch=True)
+                out = func(xor_inputs)
+                real_out = np.array(real_out)
+                out = np.array(out)
+                print(real_out, out)
+                assert np.allclose(real_out, out)
+        except AssertionError:
+            np.save("err_nodes.npy", self.pop_nodes)
+            np.save("err_connections.npy", self.pop_connections)
+
+        # print(g)
 
         self.species_controller.speciate(self.pop_nodes, self.pop_connections, self.generation)
 
         self.expand()
 
+
     def auto_run(self, fitness_func, analysis: Union[Callable, str] = "default"):
         for _ in range(self.config.neat.population.generation_limit):
             forward_func = self.ask(batch=True)
@@ -77,6 +109,7 @@ class Pipeline:
             self.tell(fitnesses)
         print("Generation limit reached!")
 
+
     def update_next_generation(self, crossover_pair: List[Union[int, Tuple[int, int]]]) -> None:
         """
         create the next generation
@@ -105,6 +138,7 @@ class Pipeline:
 
         # mutate
         new_node_keys = np.array(self.fetch_new_node_keys())
+
         m_npn, m_npc = self.mutate_func(npn, npc, new_node_keys)  # mutate_new_pop_nodes
 
         # elitism don't mutate
@@ -122,6 +156,7 @@ class Pipeline:
                 unused.append(key)
         self.new_node_keys_pool = unused + self.new_node_keys_pool
 
+
     def expand(self):
         """
         Expand the population if needed.
@@ -133,14 +168,15 @@ class Pipeline:
         pop_node_sizes = np.sum(~np.isnan(pop_node_keys), axis=1)
         max_node_size = np.max(pop_node_sizes)
         if max_node_size >= self.N:
-            print(f"expand to {self.N}!")
             self.N = int(self.N * self.expand_coe)
+            print(f"expand to {self.N}!")
             self.pop_nodes, self.pop_connections = expand(self.pop_nodes, self.pop_connections, self.N)
 
             # don't forget to expand representation genome in species
             for s in self.species_controller.species.values():
                 s.representative = expand_single(*s.representative, self.N)
 
+
     def fetch_new_node_keys(self):
         # if remain unused keys are not enough, create new keys
         if len(self.new_node_keys_pool) < self.pop_size:
@@ -153,6 +189,7 @@ class Pipeline:
         self.new_node_keys_pool = self.new_node_keys_pool[self.pop_size:]
         return res
 
+
     def default_analysis(self, fitnesses):
         max_f, min_f, mean_f, std_f = max(fitnesses), min(fitnesses), np.mean(fitnesses), np.std(fitnesses)
         species_sizes = [len(s.members) for s in self.species_controller.species.values()]
@@ -162,4 +199,4 @@ class Pipeline:
         self.generation_timestamp = new_timestamp
 
         print(f"Generation: {self.generation}",
-              f"fitness: {max_f}, {min_f}, {mean_f}, {std_f}, Species sizes: {species_sizes}, Cost time: {cost_time}")
+              f"fitness: {max_f}, {min_f}, {mean_f}, {std_f}, Species sizes: {species_sizes}, Cost time: {cost_time}")

algorithms/neat/species.py

@@ -138,7 +138,8 @@ class SpeciesController:
         for sid, s in self.species.items():
             # TODO: here use mean to measure the fitness of a species, but it may be other functions
             s.member_fitnesses = s.get_fitnesses(fitnesses)
-            s.fitness = np.mean(s.member_fitnesses)
+            # s.fitness = np.mean(s.member_fitnesses)
+            s.fitness = np.max(s.member_fitnesses)
             s.fitness_history.append(s.fitness)
             s.adjusted_fitness = None
 

examples/xor.py

@@ -17,7 +17,7 @@ def evaluate(forward_func: Callable) -> List[float]:
     :return:
     """
     outs = forward_func(xor_inputs)
-    fitnesses = np.mean((outs - xor_outputs) ** 2, axis=(1, 2))
+    fitnesses = 4 - np.sum(np.abs(outs - xor_outputs), axis=(1, 2))
     # print(fitnesses)
     return fitnesses.tolist()  # returns a list
 
@@ -38,4 +38,5 @@ def main():
 
 
 if __name__ == '__main__':
+    np.random.seed(63124326)
     main()

utils/default_config.json

@@ -2,14 +2,14 @@
     "basic": {
         "num_inputs": 2,
         "num_outputs": 1,
-        "init_maximum_nodes": 10,
-        "expands_coe": 2
+        "init_maximum_nodes": 5,
+        "expands_coe": 1.5
     },
     "neat": {
         "population": {
             "fitness_criterion": "max",
-            "fitness_threshold": 3,
-            "generation_limit": 100,
+            "fitness_threshold": 76,
+            "generation_limit": 1000,
             "pop_size": 100,
             "reset_on_extinction": "False"
         },
@@ -17,8 +17,6 @@
             "bias": {
                 "init_mean": 0.0,
                 "init_stdev": 1.0,
-                "max_value": 30.0,
-                "min_value": -30.0,
                 "mutate_power": 0.5,
                 "mutate_rate": 0.7,
                 "replace_rate": 0.1
@@ -26,34 +24,23 @@
             "response": {
                 "init_mean": 1.0,
                 "init_stdev": 0.0,
-                "max_value": 30.0,
-                "min_value": -30.0,
                 "mutate_power": 0.0,
                 "mutate_rate": 0.0,
                 "replace_rate": 0.0
             },
             "activation": {
                 "default": "sigmoid",
-                "options": "sigmoid",
+                "options": ["sigmoid"],
                 "mutate_rate": 0.01
             },
             "aggregation": {
                 "default": "sum",
-                "options": [
-                    "product",
-                    "sum",
-                    "max",
-                    "min",
-                    "median",
-                    "mean"
-                ],
+                "options": ["sum"],
                 "mutate_rate": 0.01
             },
             "weight": {
                 "init_mean": 0.0,
                 "init_stdev": 1.0,
-                "max_value": 30.0,
-                "min_value": -30.0,
                 "mutate_power": 0.5,
                 "mutate_rate": 0.8,
                 "replace_rate": 0.1
@@ -65,7 +52,6 @@
         "genome": {
             "compatibility_disjoint_coefficient": 1.0,
             "compatibility_weight_coefficient": 0.5,
-            "feedforward": "True",
             "single_structural_mutation": "False",
             "conn_add_prob": 0.5,
             "conn_delete_prob": 0.5,
@@ -81,28 +67,5 @@
             "survival_threshold": 0.2,
             "min_species_size": 1
         }
-    },
-    "hyperneat": {
-        "substrate": {
-            "type": "feedforward",
-            "layers": [
-                3,
-                10,
-                10,
-                1
-            ],
-            "x_lim": [
-                -5,
-                5
-            ],
-            "y_lim": [
-                -5,
-                5
-            ],
-            "threshold": 0.2,
-            "max_weight": 5.0
-        }
-    },
-    "es-hyperneat": {
     }
 }