Tournament [update on Dec 1, 2024]

Instructions

• Upload your agent (and related files) to this github repo (You can find your group's directory, agent_groupX.)
• There is no need to create a separate branch to upload the agent. Just push everything to the main branch.
• Right before every game, you should pull the repo so you have the latest opponent agent
  • If a group fails to upload the updated agent before the cutoff time, we will use the default one pre-uploaded to the repo
• Please designate one person as an executor for all games for the smooth transitions
  • Please make sure this person's laptop has the libraries (PyTorch, Tensorflow, SKlearn, ...)
• For group stage, every group will compete with the other groups in the same group stage. Only one group proceeds to the knockdown stage.
• If there are two groups with the same results, we will break the tie by having another game with a different reward environment.

Our Hex Game

We are going to create a common Hex Game environment OurHexGame for our PA5 and final project.

Assumption

• The board size is 11x11.

Agents

• possible_agents [“player_1”, “player_2”]
• “player_1”: red, vertical
• “player_2”: blue, horizontal

Observation

from gymnasium.spaces import Dict, Box, Discrete

Dict({
  "observation": spaces.Box(board_size, board_size),
  "pie_rule_used": spaces.Discrete(2), # 1 if used, 0 otherwise
})

• Observation Values (cell values):
  • empty: 0
  • player_1: 1
  • player_2: 2
• Use the cell ID based on the first image on https://en.wikipedia.org/wiki/Hex_(board_game) (e.g., 1A, 1B, ...)

Info

• horizontal (0) or vertical (1)
• The environment should provide an action mask to indicate invalid actions. (We can repurpose the observation.）
  • In the obs, if a hex is marked with 1 or 2, the mask should be 0. (invalid)
  • Otherwise, the mask is 1. (valid)
  • pie rule

Action

• Discrete(board_size x board_size + 1)
  • Line 1: 1A (0), 1B, ...., 1K (10)
  • Line 2: 2A (11), 1B, ...., 2K (21)
  • ....
  • the last action is the pie rule

Reward Sparse

• define sparse_flag to turn on/off the sparse reward env. If False, it should use the dense reward env.
• Win +1
• Lose -1
• Otherwise, 0

Reward Dense

• Each step, -1
• Win +floor((board_size * board_size)/2)
• Lose -ceil((board_size * board_size)/2)

Termination

• (run DFS to check the winner after each cycle) If there is a winner, terminate
• If illegal move, terminate (reset)
  • Agent should not try the illegal actions!

Rendering

please someone donate your code! thank you!

Runner

• See myrunner-eg.py

env = OurHexGame(board_size=11, sparse_flag=True) # or False
agent = GXXAgent(env)
...
action = agent.select_action(observation, reward, termination, truncation, info)