PTCG AI Battle Playground

PTCG AI Battle Challenge Playground

Status

Active — baseline validated locally, no Kaggle submission yet.

This case study develops an agent for best-of-three Pokémon Trading Card Game matches in the CABT simulator. The target is an evolving skill rating from games against other agents, followed by a final Bradley–Terry tournament—not a static prediction metric.

Competition constraints

Item	Constraint
Final submission deadline	8 January 2027, 23:59 UTC
Daily submissions	5
Active entries	Latest 2 only
Bundle	Top-level main.py and deck.csv in .tar.gz
Runtime	2 vCPU, 12.2 GiB RAM, 11.8 GiB disk
Submission limit	197.7 MiB

Baseline contract

The starter agent returns the official CABT environment's known-valid deck at initialization and then chooses the first legal actions. It is intentionally a transport and packaging baseline, not a competitive strategy. It validates a local best-of-three game and produces a correctly structured archive; no external submission has been consumed.

Development path

1. Evaluate policy changes in paired local games with seats alternated.
2. Build state/action features from card metadata and legal options.
3. Add a tactical policy for setup, energy, attacks, evolution, retreat, and prize races.
4. Search deck and policy separately before combining them.
5. Add bounded CABT look-ahead, then self-play value/policy learning.
6. Keep stable and experimental agents distinct; submission remains human-approved.

Read the versioned competition report for local commands and the evolving evidence log.