Implemented SAFE coevolution method for solving maze navigation problem

data/safe.yml

@@ -0,0 +1,98 @@
+#############################
+# The SAFE population options
+#############################
+# Probability of mutating a single trait param
+trait_param_mut_prob:  0.5
+# Power of mutation on a single trait param
+trait_mutation_power:  1.0
+# The power of a link weight mutation
+weight_mut_power:  2.5
+
+# 3 global coefficients are used to determine the formula for computing the compatibility between 2 genomes.
+# The formula is: disjoint_coeff * pdg + excess_coeff * peg + mutdiff_coeff * mdmg.
+# See the Compatibility method in the Genome class for more info
+# They can be thought of as the importance of disjoint Genes, excess Genes, and parametric difference between Genes of
+# the same function, respectively.
+disjoint_coeff:  1.0
+excess_coeff:  1.0
+mutdiff_coeff:  0.4
+
+# This global tells compatibility threshold under which two Genomes are considered the same species
+compat_threshold:  3.0
+# How much does age matter? Gives a fitness boost up to some young age (niching). If it is 1, then young species get no fitness boost.
+age_significance:  1.0
+# Percent of average fitness for survival, how many get to reproduce based on survival_thresh * pop_size
+survival_thresh:  0.2
+
+# Probabilities of a non-mating reproduction
+mutate_only_prob:  0.25
+# Probability of genome trait mutation
+mutate_random_trait_prob:  0.1
+# Probability  of link trait mutation
+mutate_link_trait_prob:  0.1
+# Probability of node trait mutation
+mutate_node_trait_prob:  0.1
+# Probability of link weight value mutation
+mutate_link_weights_prob:  0.9
+# Probability of enabling/disabling of specific link/gene
+mutate_toggle_enable_prob:  0.0
+# Probability of finding the first disabled gene and re-enabling it
+mutate_gene_reenable_prob:  0.0
+# Probability of adding new node
+mutate_add_node_prob:  0.03
+# Probability of adding new link between nodes
+mutate_add_link_prob:  0.08
+# Probability of making connections from disconnected sensors (input, bias type neurons)
+mutate_connect_sensors: 0.5
+
+# Probability of mating between different species
+interspecies_mate_rate:  0.001
+# Probability of mating this Genome with another Genome g. For every point in each Genome, where each Genome shares
+# the innovation number, the Gene is chosen randomly from either parent.  If one parent has an innovation absent in
+# the other, the baby may inherit the innovation if it is from the more fit parent.
+mate_multipoint_prob:  0.3
+# Probability of mating like in multipoint, but instead of selecting one or the other when the innovation numbers match,
+# it averages their weights.
+mate_multipoint_avg_prob:  0.3
+# Probability of mating similar to a standard single point CROSSOVER operator. Traits are averaged as in the previous two
+# mating methods. A Gene is chosen in the smaller Genome for splitting. When the Gene is reached, it is averaged with
+# the matching Gene from the larger Genome, if one exists. Then every other Gene is taken from the larger Genome.
+mate_singlepoint_prob:  0.3
+
+# Probability of mating without mutation
+mate_only_prob:  0.2
+
+# Probability of forcing selection of ONLY links that are naturally recurrent
+recur_only_prob:  0.0
+
+# The number of babies to stolen off to the champions
+babies_stolen:  0
+# The population size as a number of organisms
+pop_size:  100
+# Age when Species starts to be penalized
+dropoff_age:  50
+# Number of tries mutate_add_link will attempt to find an open link
+newlink_tries:  50
+# Tells to print population to file every n generations
+print_every:  10
+
+# The number of runs to average over in an experiment
+num_runs:  100
+# The number of epochs (generations) to execute training
+num_generations: 100
+
+# The epoch's executor type to apply [sequential, parallel]
+epoch_executor: sequential
+
+# The genome compatibility method to use [linear, fast]. The latter is best for bigger genomes
+genome_compat_method: linear
+
+# The log level
+log_level: info
+
+# The nodes activation functions list to choose from (activation function -> it's selection probability)
+node_activators:
+  - SigmoidBipolarActivation 0.25
+  - GaussianBipolarActivation 0.35
+  - LinearAbsActivation 0.15
+  - SineActivation 0.25

data/safeobjfuncstartgenes.yml

@@ -0,0 +1,27 @@
+# The objective function candidate start genome
+genome:
+  id: 1
+  # The traits used in this genome
+  traits:
+    - {id: 1,  params: [0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]}
+
+  # The neuron nodes list for this genome
+  nodes:
+    - {id: 1,  trait_id: 0, type: BIAS, activation: LinearActivation}
+    # The input sensors
+    - { id: 2,  trait_id: 0, type: INPT, activation: LinearActivation }
+    # The non-linear hidden control node
+    - { id: 3,  trait_id: 0, type: HIDN, activation: SigmoidSteepenedActivation }
+    # The output nodes
+    # The a coefficient
+    - { id: 4,  trait_id: 0, type: OUTP, activation: LinearActivation }
+    # The b coefficient
+    - { id: 5,  trait_id: 0, type: OUTP, activation: LinearActivation }
+
+  # The connection genes
+  genes:
+    - {src_id: 1,  tgt_id: 3,  weight: 0.0, trait_id: 1, innov_num: 1,  mut_num: 0, recurrent: false, enabled: true}
+    - {src_id: 2,  tgt_id: 3,  weight: 0.0, trait_id: 1, innov_num: 2,  mut_num: 0, recurrent: false, enabled: true}
+    # Hidden-to-output
+    - {src_id: 3,  tgt_id: 4,  weight: 0.0, trait_id: 1, innov_num: 12,  mut_num: 0, recurrent: false, enabled: true}
+    - {src_id: 3,  tgt_id: 5,  weight: 0.0, trait_id: 1, innov_num: 13,  mut_num: 0, recurrent: false, enabled: true}

examples/maze/common.go

@@ -168,7 +168,7 @@ func mazeSimulationStep(env *Environment, phenotype *network.Network, netDepth i
 
 	// use the net's outputs to change heading and velocity of maze agent
 	if err := env.ApplyOutputs(phenotype.Outputs[0].Activation, phenotype.Outputs[1].Activation); err != nil {
-		neat.ErrorLog("Failed to apply outputs")
+		neat.ErrorLog(fmt.Sprintf("Failed to apply outputs: %s", err))
 		return err
 	}
 
@@ -196,3 +196,9 @@ func adjustSpeciesNumber(speciesCount, epochId, adjustFrequency, numberSpeciesTa
 		}
 	}
 }
+
+// NoveltyMetric the novelty metric function for maze simulation
+var NoveltyMetric neatns.NoveltyMetric = func(x, y *neatns.NoveltyItem) float64 {
+	diff := histDiff(x.Data, y.Data)
+	return diff
+}

examples/maze/maze_ns.go

@@ -16,12 +16,6 @@ import (
 // The initial novelty threshold for Novelty Archive
 const archiveThresh = 6.0
 
-// the novelty metric function for maze simulation
-var noveltyMetric neatns.NoveltyMetric = func(x, y *neatns.NoveltyItem) float64 {
-	diff := histDiff(x.Data, y.Data)
-	return diff
-}
-
 // NewNoveltySearchEvaluator allows creating maze solving agent based on Novelty Search optimization.
 // It will use provided MazeEnv to run simulation of the maze environment. The numSpeciesTarget specifies the
 // target number of species to maintain in the population. If the number of species differ from the numSpeciesTarget it
@@ -55,7 +49,7 @@ func (e *noveltySearchEvaluator) TrialRunStarted(trial *experiment.Trial) {
 	trialSim = mazeSimResults{
 		trialID: trial.Id,
 		records: new(RecordStore),
-		archive: neatns.NewNoveltyArchive(archiveThresh, noveltyMetric, opts),
+		archive: neatns.NewNoveltyArchive(archiveThresh, NoveltyMetric, opts),
 	}
 }
 

examples/maze/maze_obj.go

@@ -43,7 +43,7 @@ func (e *objectiveEvaluator) TrialRunStarted(trial *experiment.Trial) {
 	trialSim = mazeSimResults{
 		trialID: trial.Id,
 		records: new(RecordStore),
-		archive: neatns.NewNoveltyArchive(archiveThresh, noveltyMetric, neatns.DefaultNoveltyArchiveOptions()),
+		archive: neatns.NewNoveltyArchive(archiveThresh, NoveltyMetric, neatns.DefaultNoveltyArchiveOptions()),
 	}
 }