Task 2 — Reinforcement Learning (Chef’s Hat Gym)

Student + Variant

Student ID: 16494053
16494053 mod 7 = 2
Variant: Sparse / Delayed Reward Variant

Environment

Chef’s Hat Gym is a competitive multi-agent, turn-based card-game environment with:

multi-agent non-stationarity (opponents affect dynamics)
large discrete action space
delayed / sparse rewards (match outcome)
stochasticity and sequential decisions

Agent Design (What I implemented)

Algorithm: REINFORCE (Policy Gradient) with action masking.

State Representation

A compact numeric state vector built from:

hand (normalized by 13)
board (normalized by 13) Concatenated and padded/truncated to a fixed obs_dim=28.

Action Handling Strategy

Chef’s Hat provides possible_actions.
I create an action mask over the full action list and sample only from legal actions by setting illegal logits to a very negative value.

Reward Usage (Sparse/Delayed focus)

Two experiments:

baseline_terminal_only: only terminal reward at match end (delayed credit assignment).
shaped_terminal_plus_aux: terminal reward + small auxiliary shaping when cards_left decreases.

Terminal reward mapping (from finishing position):

1st: +1.0, 2nd: +0.3, 3rd: -0.3, 4th: -1.0

How to Run

Install

pip install -r Requirements.txt
pip install -e .
pip install torch numpy pandas matplotlib

License

atluris5/TASK-2-7043-SCN

About

Resources

License

Stars

Watchers

Forks

Releases

Languages