Skip to content

Latest commit

 

History

History
99 lines (78 loc) · 3.05 KB

README.md

File metadata and controls

99 lines (78 loc) · 3.05 KB
Raw

Microverse

Build Status

Self reinforcement training of neural network using genetic algorithms.

About

In this repo we implement neural network model that is being trained using heuristics from genetic algorithms. The microverse module has simulator of 2D world in which agents can see the food they need to eat in order to stay alive. The fit function that the model is trying to optimize is the number of eaten food items. Every time an agent dies, we get the two best agents in the current population and we cross their behavior functions (the neural networks) and we mutate them - making a new agent.

Content

  • Simulator

    • Intersection algorithm
    • Set of render-able and updatable agents

  • Renderer - we use the ikinter library to visualize the running simulation

  • Agent - base class for agent - (contains health, color, position and velocity properties)

  • Plugins

    • Sight - sensor module for the agent
    • Digestion - module for consuming the food and leveling_up the agent
    • Mobility - module used for adding the velocity to the position of the agent
    • Neural network - simple implementation of neural network (it has only feed-forward and crossover as public methods) used as the brain of the agent
    • Navigator - module that consumes the output of the network and steers the agent accordingly
    • Fatigue - this module reduces the health of each agent in every tick of the simulation

Heuristics used in the genetic algorithm

Fitness function

Number of eaten food items.

Crossover

For biases

parent_a = [a1, a2, a3, a4, a5]
parent_b = [b1, b2, b3, b4, b5]
random_split = random(0, 5) # ex. 2

child = [a1, a2, b3, b4, b5]

For weights

The heuristic for weights applied for each row of the weights matrix.

Mutation

Apply gene change with probability (mutation_rate).

mutated_dna = [gene if prob() < mutation_rate else
               random(0, 1) for gene in dna]

Selection

When the target population size is more than the current - crossover the two most fit agents in the current population.

Examples

How to run a simulation

while RENDERER.is_running:
    food_spawner()
    smart_agent_spawner()

    RENDERER.update()

    for agent in SMART_AGENTS:
        agent.update()
    for food in FOODS:
        food.update()

    for agent in SMART_AGENTS:
        agent.render(RENDERER)
    for food in FOODS:
        food.render(RENDERER)

    recycle(FOODS)
    recycle(SMART_AGENTS)

How the plugin design implements the smart agent

self.sight = Sight(
    fov=fov,
    ray_count=ray_count,
    strength=strength,
    environment=environment
)
self.brain = NeuralNetwork([self.sight.ray_count] + nn + [3])
self.parents = parents
navigator = Navigator()
digestion = Digestion(environment)
mobility = Mobility()
fatigue = Fatigue()

self.plug(self.sight, self.brain, navigator)
self.plug(digestion)
self.plug(mobility)
self.plug(fatigue)