Skip to content

A 2D battleground simulation where AI's fight each other and evolve over time

License

Notifications You must be signed in to change notification settings

timjrobinson/evolutionary-ai-battle

Repository files navigation

Evolutionary AI Battle

A 2D battleground simulation where AI's fight each other and evolve over time.

Example of AI evolving over generations

Overview

Many universities and companies hold virtual AI battles for their engineers. You are given a sandbox environment with a few simple rules and must produce an AI algorithm that can beat all others. For a battle this year I decided to build an evolutionary algorithm based off the Evolving Neural Networks through Augmenting Topologies (NEAT) paper and inspired by the Mar/IO project. This required building a simulator and system to train these AI's against each other, which is what this package contains.

The system initially creates 100 random genomes, then makes them fight each other, each AI playing 5 battles per round. At the end of each round the weakest bots are culled, and the strongest bots go on to have children, which inherit their parents genes plus get some random mutations of their own. The remaining parents and all the children then play each other in the next round. Over time this leads to more and more intelligent AI's battling each other and constantly creating new strategies and counter-strategies.

Setup

npm install

The code is written using ES6 functionality and is compiled for the browser using Webpack and compiled for NodeJS using Babel. The browser entrypoint is src/index.js and the NodeJS entrypoint is src/coordinator.js.

Usage

Training

You can either train and watch one game at a time in your browser, or run many games in parallel using NodeJS. NodeJS will get you semi-intelligent AI's much quicker however you won't be able to see the games being played.

You can first train many games using NodeJS and then view some of the battles between the advanced AI's in your browser. To do this see the "Watching" section below.

Browser Training

  • npm start to compile all the files and start the webserver.
  • Open http://localhost:1337 in your browser.

You should see the red bot chilling out not doing much, while the blue bot does something randomly. In the initial stages of training only the red bot is trained, while the blue bot has a random move / shoot algorithm to give some variety to the battles. Because it's the very beginning the red bot doesn't do much because it doesn't know what to do yet, this will change after a few generations.

Headless Training

This will run all battles for each generation in parallel in headless mode. This will train 10 - 100x faster than Browser Training.

npm run train

This will compile all the src files into dist/nodejs/ then run node dist/nodejs/coordinator.js to start the training.

If training gets interrupted and you wish to start it from where it left off, there is a line at the top of src/coordinator.js which you can uncomment to resume training from that point.

If you want to see debug logs while training run training with the environment variable DEBUG set to true.

DEBUG=true npm run train

Watching

To watch a battle between two trained AI's you can do the following:

  • npm start to compile all the files and start the webserver.
  • Open http://localhost:1337 in your browser.
  • Select the name of the species you trained in Headless training. The list is sorted by date, so it's probably the first item.

Battle Statistics

In the watch mode you'll see there are battle statistics in the bottom right. These are:

  • Generation - The current generation number you are watching. Bots play in generations and each genome plays 5 battles per generation.
  • MaxFitness - The fitness of the most fit species in this generation.

Competitor Details

  • Previous Fitness - The fitness of this genome in the last generation (total fitness over all 5 games added up). New means this genome is a child that was just created and so didn't battle last generation.
  • Current Fitness - The total fitness the genome has gained in battles in this generation. Starts at 0 and after each battle the results of that battle are used to calculate the genome fitness and add it to this number. See the "How is bot fitness calculated?" section of this README for more information.

Configuration Options

All config options are stored in the file config/default.json

Species related settings

  • initialSpecies - int - The total species created at the beginning
  • initialGenomesPerSpecies - int - The total genomes in each species
  • population - int - The approximate number of genomes in each generation

Battlefield related settings

  • mapWidth - int - The width of the battlefield
  • mapHeight - int - The height of the battlefield
  • botStartPoses - object - the position and rotation of each bot at the beginning of each battle
  • startingLives - int - How many lives each bot has
  • tickTime - int - The time between each game update tick in ms
  • botSize - int - The diameter of each bot in pixels
  • maxSpeed - int - The number of pixels each bot can move each tickTime
  • bulletSize - int - The diameter of each bullet
  • bulletSpeed - int - The number of pixels each bullet moves each tickTime

Battle timeout related settings

  • maxRoundTime - int - The battle always ends after this many seconds
  • noActionTimeout - int - If the AI bot does not perform an action in this much time the battle will end
  • noMoveTimeout - int - If the AI bot does not move in this much time the battle will end

Genome related settings

  • neuralNetworkSquareSize - int - The neural network brain grid is divided into squares of this size. When a bot or bullet moves into that square it marks the square as -1 for a bullet or 1 for a bot. A lower square size will improve the fidelity of a bot, and possibly cause it to perform better, but will require more processing power.
  • maxNeurons - int - The maximum number of neurons in a bots brain. Increasing this number will allow for more complex bots but will require more processing power.

Genome mutation options

These are just default values, each genome will randomly modify these options over time

  • mutateConnectionChance - float - The chance that the weight of a connection between two neurons will be changed in mutation

  • pertubeChance - float - The chance that the weight of a connection will be scaled rather than set to a random value

  • mutateLinkChance - float - The chance that a new link will be created between two nodes

  • mutateNodeChance - float -The chance that a new node will be placed on a link, creating to new links to the previous input and output nodes of that link

  • mutateDisableChance - float - The chance that a link will be randomly disabled

  • mutateEnableChance - float - The chance that a link will be randomly enabled

  • stepSize - float - When pertubing a weight the amount that the weight will be changed by.

  • secondParentInnovationGeneChance - float - When having children most genes come from the fittest parent, this is the chance that genes with the same innovation number will come from the less fit parent.

  • disjointMultiplier - float - When determining if a genome is part of a certain species this is a multiplier of the number of genes in common that the bot must pass. Decreasing this number will create more unique species, increasing it will place more bots into the same species

  • weightMultiplier - float - When determining if a genome is part of a certain species this is a multiplier of the difference in weights between the genes of the genome and the species. Decreasing this number will create more unique species, increasing it will place more bots into the same species

  • deltaThreshold - float - When determining if a genome is part of a certain species this is the delta of that both the weights and joints must pass to be considered part of the species.

Definitions

  • Species - A species is a distinct set of AI's that inherit genes from each other. Each species contains many genomes.
  • Genome - A genome is a collection of genes which define how an bot acts. Each bot has one genome.
  • Gene - A gene describes a link between two neurons in the AI's brain. Each gene belongs to one genome and a genome has many genes.
  • Innovation - An innovation number is given to each gene upon creation. If a gene is passed to a child it is given the same innovation number. When two parents create children all genes come from the strongest parent, except those genes with the same innovation number, they are picked from one parent at random.

Implementation FAQ

How do battles work?

The battle is played in 75ms ticks. Each tick each AI is given the following information:

  • xPos - int [0 - MAP_WIDTH] Your bots x (horizontal) coordinate, where 0 is the left hand most column and MAP_WIDTH the right hand most column.
  • yPos - int [0 - MAP_HEIGHT] - Your bots y (vertical) coordinate, where 0 is the top of the screen and MAP_HEIGHT is the bottom.
  • rotation - int [0 - 360] - The angle your bot is facing in degrees, where 0 degrees is east, 90 degrees is north, 180 west, 270 south.
  • bullets - Array - Any of your bullets currently on the screen. Each bullet is an object of {xPos, yPos, rotation}. otherPlayer - An object with the above properties for the other player

This object is sent to the bot each tick and the bot uses that to make it's next decision. Then it must return an object containing the following:

  • ds - boolean - Is the bot shooting or not
  • dx - int [-15 - 15] - The x-speed of the bot, from -15 (west) to 15 (east)
  • dy - int [-15 - 15] - The y-speed of the bot, from -15 (north) to 15 (south)
  • dh - int [-15 - 15] - The rotation speed of the bot, where -15 is counterclockwise and 15 is clockwise

You can only have a maximum of 5 bullets on the screen at once and they are destroyed when they hit the opponent or wall. Each player has 5 lives.

How is bot fitness calculated?

Bot fitness is used to determine what Species go on to produce children vs those that die out. After each battle the fitness is calculated for Bot 1 with the following calculation:

First the fitness is initialized to the generation number of the species, this gives genomes a little boost to account for battles being harder as time goes on. A bot with 100 fitness in generation 500 is probably much better than one with 100 fitness in generation 5.

The bot then gets +20 fitness for each hit on the opponent.

If the bot lost the battle it gets +1 fitness for each second it survived.

If the bot won the battle it gains:

  • +1 fitness for each second it had left until maxRoundTime (60 seconds by default), to reward it for winning quickly.
  • +10 fitness for each life it had left.
  • +100 fitness for winning.

Only Bot1 gets a fitness score and contributes to the genome/species overall fitness. Bot 2 is simply a random opponent to fight against and is discarded at the end of the battle.

Bots play 5 battles (versus a random genome each time) in each Round. The fitness from each battle is summed to get the overall fitness of the genome for species culling/reproduction.

How does the bot brain map work?

When the bot receives input from the battleground it takes those player and bullet positions and turns it into a map that is relative to the bots position and direction, so the bot is the center of the world and everything else moves/rotates around it.

This is so that the bot can learn quicker, as it can learn that if a bullet is heading towards it. from the right it needs to move forward or backward to dodge it etc.

If the bot neural network just took in the raw data it would take a lot longer to learn as the bots decisions would be highly related to it's current position and it would be almost impossible for it to come up with a general algorithm to dodge bullets or aim at the enemy in every possible position/rotation combination.

What are Innovation Numbers?

The reason genes have innovation numbers is because if you picked genes randomly the children of even smart parents would be pretty stupid.

For example you could have a parent that is good at shooting and a parent that is good at moving, and if they have a child you'd expect it would take one 'shooting' gene, and one 'moving' gene from each parent. If you picked genes randomly the child may instead get a 'moving' gene from each parent and no gene for shooting, and it would be significantly worse than the previous generation.

By giving an innovation number to each gene, and genes keep that innovation number when passing to children, you can make children only swap out genes that have the same genesis, so they are likely to be doing roughtly the same thing. This way all children of the above parents should have at least one shooting gene, and one moving gene, and if they're lucky they'll get the best genes from each parent (but that part is random, it is evolution after all).

License

MIT © Tim Robinson

About

A 2D battleground simulation where AI's fight each other and evolve over time

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published