Developing an Agent

In order for the game to properly load your agent one must install the agent, there are several ways to do so: 1. create a new agent within src/splendor/agents and when installing splendor your agent will be installed as well. (i.e. when invoking pip install .) 2. create a new package and develop your agent there and then install it. 3. create a new agent within src/splendor/agents and ONLY DURING DEVELOPMENT install splendor by using pip install -e . (instead of the pip install .) which allowes you to edit and adjust your agent as you please without the necessity to re-install the package.

Training Our Agents:

Training The Genetic Algorithm Agent:

In order to train the genetic algorithm agent with the following hyper-parameters: 1. Specify the population size in each generation to be 24 (should be a multiple of 12). 2. Train for 20 generations. 3. Fix the mutation rate chance to be 0.1(%). 4. Use a fixed random seed. Use the following command:

evolve --population-size 24 --generations 20 --mutation-rate 0.1 --seed 1234

Training The PPO Agent:

In order to train the PPO agent you should run the following command:

ppo

This command will train the PPO agent with the default training hyper-parameters.

ppo --device cuda --working-dir runs --transfer-learning --opponent minimax

This command will use GPU during it’s training, it will use the installed weights as initialization of the network and the PPO will be trained agaisnt MiniMax. Furthermore all the generated files (weights stored in .pth files and stats.csv) will be generated within the directory runs/.

There are multiple available architectures for the neural network to be used by the PPO agent, for instance:

  1. MLP - Multi-Layered Perceptron, also known as Fully-Connected Feed-Forward network.

  2. Self-Attention and then MLP.

  3. GRU (Gated Recurrent Unit) and then MLP.

  4. LSTM (Long Short-Term Memory) and then MLP.

By default the MLP architecture will be used, however you can decide to train the PPO with a different architecture via the --architecture or via it’s shortcuts -a and --arch, here is an example:

ppo --device cpu --opponent random --architecture gru

There are also multiple opponents available to be trained or evaluated agaisnt, such as:

  1. random

  2. minimax

  3. itself

  4. PPO (with MLP)

  5. PPO (with Self-Attention)

  6. PPO (with GRU)

  7. PPO (with LSTM)

SplendorEnv - an OpenAI gym compatible simulator for the game Splendor

We’ve made a custom gym.Env and registered it as one of gym environments. This would come in handy when training agent such as DQN or PPO.

How to create an instance of SplendorEnv:

  1. import gymnasium - import gymnasium as gym.

  2. registering SplendorEnv to gym - import splendor.Splendor.gym

  3. define the opponents:

When creating an instance of SplendorEnv you should tell it which agents will be used as opponents to you (the one who uses the env.). For the following example we’ll use a single random agent as an opponent.

from splendor.agents.generic.random import myAgent

opponents = [myAgent(0)]

  1. creating the environment:

env = gym.make("splendor-v1", agents=opponents)

Custom features of SplendorEnv

  1. every call to env.step(action) simulate (by using SplendorGameRule) the turns of all the opponents.

  2. when calling env.reset() SplendorEnv will return the feature vector of the initial state AND the turn of our agent via the second variable (the dict) which will have a key called my_turn.

  3. SplendorEnv have several custom properties:

    1. state - the actual SplendorState - not the feature vector.

    2. my_turn - the turn of the agent, same as the value returned by env.reset().

  4. SplendorEnv have several custom methods:

    1. get_legal_actions_mask - a method for getting a mask vector which masks all the illegal action of splendor.Splendor.gym.envs.actions.ALL_ACTIONS.

You can access those like this:

env.unwrapped.my_turn