rahulbhatia / splendor-ai

Splendor AI Project

Purpose

• Model a "real-world" system
• Be able to create a complex state machine
• Model interactions with that state machine in a simple to understand way
• Build an ai solution that can make a decision on which move to play
• Feature engineering on how to transform the game model into inputs to a statistical model

Part 1: Modeling

• Model the Splendor Game by decomposition into Decks, Table, Banks, & Players
• Model the Players as Banks, Cards, + Points
• Model the Banks as dictionaries of Stacks (which are just unit counters)
• Model the Table as face up set of cards
• Be able to generate a realistic looking deck for either l1,l2,l3 card types

Part 2: Model State Change

• Implement state change for each model type
• Decks generate cards
• Stacks can grow or decrease
• Players can do one of a set of actions {take 2 coins of one color, take 3 different coins, buy a card}
• Be able to apply a state change to a game and generate the resulting game object
  • by this approach, a SplendorGame object is immutable; rather than mutate the state, we actually produce a new SplendorGame by applying a SplendorMove(or something similar)
  • We do this because our eventual ai system will need to apply all possible moves and evaluate the best one using some heuristic (trained or otherwise)

Part 3: Simulate Game Play

• Construct a notation for gameplay
• simulate gameplay as a random walk through space

Part 4: Build a Policy Model to evaluate the $P(Player_{x} Wins \mid State )$

• Generate a few potential models to evaluate the probability above
• Have the models play each other to see who wins the most