FlappyBird AI

The cover image shows the “brain” — read: neural network — of a bird that I designed, and how it chooses when to flap to complete an obstacle course.

The animation above shows the learning process. By randomly creating a batch of different brains, having them complete the course, and then creating a new generation based on the birds that went the farthest, we can “evolve” a bird that is better at the game.

Development

The point of this project was to make a computer learn how to play a game effectively. I developed a neural network in C++ that evolved using a genetic algorithm. The “fittest” birds reproduce the most, and those are the ones that make it through the most obstacles. The brain works by feeding in several inputs and producing an output that tells the bird to flap or not. After some experimentation I found that one layer of six nodes worked best for producing a bird that learned quickly.

How neural networks and genetic algorithms fit together

Put simply, neural networks are a way of approximating how neurons function in your brain. Each neuron sums the values of the neurons before it according to a given weight, and the last neuron is the output neuron that indicates when the bird should flap. Each neuron’s weight represents how much the brain cares about a specific input — speed, distance to the top of the obstacle, and so on — which lets the brain make complex decisions.

These weights are modified after all the birds have failed. The one that moved farthest is the most likely to reproduce, following the biological process in nature. Two bird brains are chosen and their “chromosomes” — the associated weights in the network — are crossed over to form a child. That child also has a chance of a slight mutation, which nudges a weight.