AIsteroids - Asteroids AI
This project is actually a combination of two smaller projects I wanted to do. One is to build a game server which used websockets, and the other an artificial intelligence (AI) using neural networks. The desire to build the game server came from the fact that I had previously built other retro games but they were all frontend, it would be nice to have a client to send commands to a server so that people could play together (and prevents cheating to some degree). The desire to build a neural network actually came from a baby book given to me by a friend which was titled 'My first neural network', which is a pretty picture book to explain to baby's what a neural network is - it's a bit of a jokey present, but the book is very cute and made me realise that I should make a neural network to have a better foundation in AI.
The Asteroid Game
So the game is a simple asteroids game. You control a ship which can rotate and thrust forward. You can shoot bullets to destroy asteroids. The asteroids break into smaller asteroids when shot. The game is over when you collide with an asteroid.
The actual asteroid game includes space aliens which will try to shoot at you, but this was a minimal viable product and a working game.
So the implementation was a lot more straight forward than I thought it would be. Theres a 'Ship' object, 'Asteroid' objects, a 'Project' object and a 'Game' object.
Because I had no idea how fast the neural network was going to be, the game will only progress to the next 'frame' if you send it your current actions. So the client will initialise a game to the server. The server will respond with the current ship, asteroids and projectiles positions and sizes. The server won't load the next 'frame' until the client has responded with their actions i.e. fire, thrust, turn left or turn right.
The server will then calculate the next frame and send it back to the client. The client will then render the new frame.
This loop is essentially the game
The server also sends the locations of the ship, asteroids and projectiles against a timestamp to the database, so I can load back how each artificial brain played.
The Neural Network
The neural network is a simple feed forward neural network. It has an input layer, a hidden layer and an output layer. The input layer is the current state of the game. The output layer is the actions the ship can take. The hidden layer is the 'brain' of the neural network.
The neural network input layer consists of 'distance sensors' on the ship. Each sensor at 45 degree positions to the ship. The closer something is to the ship at that angle, the more that neuron fires. This is similar to how neuron encode things to our brains. Something warm might have a low frequency but when its 'hot' the rate at which that sensor fires is very quick.
The neural network starts with zero knowledge. It has 8 inputs, one for each sensor, and 4 outputs, each corresponding to an input to the asteroid server.
Each neural network then has random mutation(s) of the following:
- Add neuron
- Remove neuron
- Change neuron sensitivity
- Add Connection
- Remove connection
- Change connection sensistivity
- Change output
- Change 'neurotransmitter'
The neurons will fire when a threshold is met. The lower the sensitivity the more that neuron with fire. The connection sensitivity is just how much of that fired potential is actually received by the other end.
One thing I learnt when studying neurons is about all the different types of neurotransmitters. Some promote and other inhibit. And different neurons have different sensitivities for different neurotransmitters. But in my model, there is only a range between -1 (red) and 1 (green) so that some neurons can act as 'nots'
For instance, there might be an asteroid about to hit from the side, so some neurons might say to go forward. But the front sensor says there is also a neurotransmitter there. So one tactic the brain might develop is a negative output to thruster if there is an asteroid dead ahead.
So the neural network will be created with random mutation(s). The neural network will then be send create a websocket to the game server and play a game. The neural network will then be scored based on how well it played the game. The neural network with the highest score out of its siblings will be kept and the others will be discarded.
The neural network will then be mutated again and the process will repeat.
So the neural network will evolve to play the game better and better.
I ran the simulation from the beginning multiple times and noticed two main patterns. One where it will try to zip around and shoot asteroids. The other is it just stays still and shoots at the asteroids.
I've attached a gif of one of the more interesting games I saw the NN (neural network) play below.
I've also attached a gif of how the brain grows with each generation (note it isn't the same neural network which plays the asteroid game on the left. This brain developed the strategy to remain still! Which can be seen from the lack of connections to the thruster. But the lack of movement didn't make a good video!)
Future of this project
There are a few things I would like to do to improve the game and AI.
The next thing I would do is implement the aliens/UFO which target the players ship. This would stop the AI developing the tactic to remain still
Another thing I would like to do is make the game and AI more performant. The AI is written ts for a node server, which only uses one of the cores and doesn't offload the repetitive mathematics to the GPU, both of which would increase the speed, and therefore would progress through the generations a lot faster.
With more performance, I can then make the brain more complex - like coding more neurotransmitters to see how the brain would react, i.e. would we see different neurotransmitters fire off for specific purposes like when best to shoot or when best to flee.
Finally I would like to tweak the logic to a recurrent neural network where information loops back.