lander.py

# Solution organism was found in 78th generation
# Total of 200 * 5 * 78 = 78000 trials were used for training


from __future__ import print_function

import gym
import numpy as np
import itertools
import os

from neat import nn, population, statistics

np.set_printoptions(threshold=np.inf)
env = gym.make('LunarLanderContinuous-v2')

# run through the population


def eval_fitness(genomes):
    for g in genomes:

        net = nn.create_feed_forward_phenotype(g)
        total_fitness = 300

        for k in range(5):

            observation = env.reset()
            # env.render()
            fitness = 0
            reward = 0
            frames = 0

            while 1:
                inputs = observation
                # print(inputs)
                inputs = np.append(inputs, np.array(reward))
                # print (inputs)
                # active neurons
                output = net.serial_activate(inputs)

                output = np.clip(output, -1, 1)
                # print(output)
                observation, reward, done, info = env.step(np.array(output))


                fitness += reward
                frames += 1
                # env.render()
                if done or frames > 1000:
                    if fitness < total_fitness:
                        total_fitness = fitness
                    # print(fitness)

                    break
        # evaluate the fitness
        # fitness is lowest during trial
        g.fitness = total_fitness
        print(g.fitness)

local_dir = os.path.dirname(__file__)
config_path = os.path.join(local_dir, 'lander_config')

pop = population.Population(config_path)
pop.run(eval_fitness, 1000)
winner = pop.statistics.best_genome()
del pop

winningnet = nn.create_feed_forward_phenotype(winner)

env.monitor.start('walker-experiment/', force=True)


streak = 0



while streak < 100:
    fitness = 0
    frames = 0
    reward = 0
    observation = env.reset()
    env.render()
    while 1:
        inputs = observation

        inputs = np.append(inputs, np.array(reward))
        # active neurons
        output = winningnet.serial_activate(inputs)
        output = np.clip(output, -1, 1)
        # print(output)
        observation, reward, done, info = env.step(np.array(output))

        fitness += reward

        env.render()
        frames += 1

        if done or frames > 1000:
            if fitness >= 200:
                    print(fitness)
                    print ('streak: ', streak)
                    streak += 1
            else:
                print(fitness)
                print('streak: ', streak)
                streak = 0

            break
print("completed!")
env.monitor.close()

lander_config

#--- parameters for the lander continuous v2---#

# The `Types` section specifies which classes should be used for various
# tasks in the NEAT algorithm.  If you use a non-default class here, you
# must register it with your Config instance before loading the config file.
[Types]
stagnation_type      = DefaultStagnation
reproduction_type    = DefaultReproduction

[phenotype]
input_nodes          = 9
hidden_nodes         = 0
output_nodes         = 2
initial_connection   = unconnected
max_weight           = 10
min_weight           = -10
feedforward          = 0
activation_functions = tanh sigmoid relu identity
weight_stdev         = 3

[genetic]
pop_size                = 200
max_fitness_threshold   = 210
prob_add_conn           = 0.3
prob_add_node           = 0.1
prob_delete_conn        = 0.05
prob_delete_node        = 0.03
prob_mutate_bias        = 0.00109
bias_mutation_power     = 0.01
prob_mutate_response    = 0.01
response_mutation_power = 0.01
prob_mutate_weight      = 0.3
prob_replace_weight     = 0.03
weight_mutation_power   = 0.1
prob_mutate_activation  = 0.01
prob_toggle_link        = 0.0138
reset_on_extinction     = 1

[genotype compatibility]
compatibility_threshold = 3
excess_coefficient      = 1.0
disjoint_coefficient    = 1.0
weight_coefficient      = 0.4

[DefaultStagnation]
species_fitness_func = mean
max_stagnation       = 5

[DefaultReproduction]
elitism              = 3
survival_threshold   = 0.2

To be clear: you uploaded to the open AI gym only the result of your test runs and not your training?

Yes Madcow.

The reason for doing this is because NEAT isn't really a training algorithm. It's more of a search algorithm, so showing every evaluation will create charts that don't make a lot of sense without knowing which species/organism is being evaluated.