pokemon-learning

Practicing reinforcement learning techniques using Pokemon battle rules.

##Goals Our goal is to analyze Pokemon stats and movesets to form the ultimate team, then teach that team to win using reinforcement learning.

##Prerequisites The data folder must contain the following files: moves.json, pokemon-listing.txt, and pokemon-type-listing.txt. There also should be a PokemonToAnalyze file with a newline-separated list of Pokemon to generate data for.

##Building and Running Use ./gradlew build to build the app. This compiles the code and runs all the tests.

####Data Generator To run the data generation, use the following command: ./gradlew runDataGenerator. To run it with arguments, add -Dexec.args="<arguments here>" to the end of that command. Note that it will look at the data/PokemonToAnalyze for which Pokemon to run the data on. The output will be a folder named output with a teams.json file and a folder named simulated-battles containing an indexed collection of folders of all the team battles generated. Each file will contain the collection of battles for a given team as T1.

######Arguments EXAMPLE: -Dexec.args="-setSize 2 -numberOfBattles 1- -skipExistingFalse false"

• -setSize defaults to 3.
  • Size of sets of Pokemon to analyze.
  • For example, if the setSize is 2, we generate battles for all teams that contain at least 2 team members from the PokemonToAnalyze set.
• -numberOfBattlesdefaults to 5.
  • Number of battles to simulate per team having battles generated.
  • Note that the higher this is, the longer this takes to run.
• -skipExistingFiles defaults to true
  • True means that we do not generate battles if a file for the team already exists
  • Used if the process was killed prematurely upon initial generation of data and needs to be restarted.
  • False means that it reads in the battles generated and adds onto the existing list with more battles

####Pokemon Performance Analysis To run the Pokemon Analysis, use the following command: ./gradlew runPokemonAnalysis 1. Note that the number at the end represents the number of Pokemon to look at in combination. For example, if it's 2, we look at each pair of Pokemon in the PokemonToAnalyze file and calculate the win percentages for teams containing that pair. If it's 1, we look at win percentages for teams that contain one of the singular Pokemon in the file. If it's 3, we look at the winning percentage of for just that team of 3. All output goes to the console.

####Pokemon Move Analysis To run the Pokemon Move Analysis, use the following command: ./gradlew runMoveAnalysis. This takes in no arguments. It creates a file in the output folder called move-analysis.json that contains a list of JSON objects with a mapping of Pokemon name to a mapping of moves and counts of times the move was used among all battles with that Pokemon. Note that if you quit the process before it is finished, it writes to file everything it had currently calculated.

####Pokemon Next Move Analysis To run the Pokemon Next Move Analysis, use the following command: ./gradlew runNextMoveAnalysis. This looks at each Pokemon in the PokemonToAnalyze file and determines which move it should use against every possible defender and the confidence that it will lead to a team win. The output goes to the output folder called pokemon-next-moves.json, which contains a list of objects with an attacker, defender, move, and confidence. Like the move analysis, this writes to file everything it had calculated when quit prematurely.

##Notes The initial PokemonToAnalyze file before generating data should contain the following pokemon for best results:

Mewtwo
Dragonite
Mew
Moltres
Zapdos
Articuno
Cloyster
Gyarados
Arcanine
Exeggutor
Tauros
Lapras
Rhydon
Aerodactyl
Tentacruel
Starmie
Pinsir
Vaporeon
Jolteon
Flareon
Kabutops
Snorlax
Venusaur
Charizard
Blastoise
Gengar
Kingler
Omastar
Ninetales
Machamp
Victreebel
Golem
Rapidash
Weezing
Scyther
Chansey
Nidoqueen
Nidoking
Poliwrath
Hypno
Kangaskhan
Sandslash
Golduck
Alakazam
Dewgong
Dodrio
Muk
Electrode
Pidgeot
Raichu

All data of Pokemon's moves and types came from pokemondb.net. The tasks will take in the numberhood of hours to complete, so be careful with your inputs.