dgisim 0.3.dev0

A highly customizable Genius Invokation TCG Simulator for AI training

Dottore Genius Invokation TCG Simulator

A Genshin Impact Genius Invokation TCG simulator intended to be used for AI training.

• Documentation
• Source code
• Contributing
• Game design
• Bug reports
• Discord server

This package aims to help programmers code things based on Genius Invokation TCG with ease. e.g. AI, desktop application, website...

The simulator is modeled as a finite state machine, where all game states are immutable. Optimizations are done to make sure immutability doesn't impact performance.

Basic rules of Genius Invokation TCG can be found on Fandom.

Installation

Please make sure your Python version >= 3.10 before installing.

pip install dgisim

Simple Start With CLI

Run Locally

Once installed, you may start by trying the CLI to play the game first.

You might want to run a simple python program like this:

from dgisim import CLISession

session = CLISession()
session.run()

Run Remotely

You may try the CLI online on Google Colab

CLI Simple Usages

See CLI's README for showcase and explanations of the CLI.

Customize Player Agents (Important For AI Or Building App)

A player agent controls all actions of a player in a game.

To implement a player agent, all you need to do is to inherit the abstact class PlayerAgent and implement the method choose_action().

A simple example is shown below, the agent implemented choose 3 random cards to replace during Card Select Phase, and normal attacks until there's no dices for it during Action Phase.

class ExampleAgent(PlayerAgent):
    def choose_action(self, history: list[GameState], pid: Pid) -> PlayerAction:
        latest_game_state: GameState = history[-1]
        game_mode: Mode = latest_game_state.get_mode()
        curr_phase: Phase = latest_game_state.get_phase()

        if isinstance(curr_phase, game_mode.card_select_phase):
            cards_to_select_from: Cards = latest_game_state.get_player(pid).get_hand_cards()
            _, selected_cards = cards_to_select_from.pick_random_cards(num=3)
            return CardsSelectAction(selected_cards=selected_cards)

        elif isinstance(curr_phase, game_mode.action_phase):
            me: PlayerState = latest_game_state.get_player(pid)
            active_character: Character = me.just_get_active_character()
            dices: ActualDices = me.get_dices()
            # check if dices are enough for normal attack
            normal_attack_cost = active_character.skill_cost(CharacterSkill.NORMAL_ATTACK)
            dices_to_use = dices.basically_satisfy(normal_attack_cost)
            if dices_to_use is not None:
                # normal attack if dices can be found to pay for normal attack
                return SkillAction(
                    skill=CharacterSkill.NORMAL_ATTACK,
                    instruction=DiceOnlyInstruction(dices=dices_to_use),
                )
            return EndRoundAction()  # end round otherwise

        else:
            raise NotImplementedError(f"actions for {curr_phase} not defined yet")

The above example manually tests if there are dices for some action, which is straightforward but takes time to exhaust all options. So the GameState can return an ActionGenerator object which automatically provides you with all valid actions to choose from. More about ActionGenerator will be updated later.

You can find more examples of implementations of PlayerAgent in dgisim/src/agents.py. The RandomAgent in agents.py is implemented based on ActionGenerator mentioned above to make random but valid decision.

Features

This simulator is modeled as a finite state machine, which means any intermediate state can be standalone and be used to proceed to other states.

The GameState class represents some game state in the state machine. It uses passed in Phase object to determine how to transform to another state, which means the game flow is highly customizable. (Default Mode and some Heated Battle Modes are implemented already)

Everything in the GameState object are immutable, so traversing game history and exploring different branches of possibilities in the future are not error-prone. The simulator did optimizations for immutability. The unchanged data are shared among neighbouring game states.

GameState implements __eq__ and __hash__, enabling you to use any game state as a key in a dictionary, and discover game states on different 'game branches' being actually the same.

An ActionGenerator can be returned by any valid GameState to help generate valid player actions.

Development Milestones

Currently a full game can be played with any combination of the characters and cards implemented.

• Implement all game phases (Action Phase, End Phase...)
• Implement all cards (34/192 implemented) (details)
• Implement all characters with their talent cards (10/51 implemented) (details)
• Implement all reactions
• Implement death handling
• Implement all game logics to support the implemented cards and characters
• Implement interactive CLI for better debugging experience
• Implement lazy player agent for minimal testing purposes
• Implement random player agent for testing purposes
• Implement player action validity checker
• Implement player action choices provider

Future Plans

I have the plan to implement a simple cross-platform GUI interface for the simulator. But that will be in a separate repo.

Once this project is done, I'll be reading relative papers and develop an AI for this game. The AI is supposed to be used for learning strategies and making decks, but not against another player directly.