The competition has finally come to an end After a week of seclusion, I only won a silver medal in the end. (The match will continue to pair for another two weeks, so if nothing unexpected happens, you should still be able to hold on to your cards.) I have created an online duelist that can compete with my AI. It will be open for a few days and everyone is welcome to come and compete with my AI! (Link in message) It's been a long time since I last played Kaggle, and surprisingly, I really enjoyed it. This AI battle is not just about ordinary agents doing harness or LLM training content, but really involves RL/model structure. After having Coding Agents in the competition, it felt very different. Although I participated alone, it was like a whole team. I almost used all the resources of the company and opened 6 Agents, 2x GROK 4.6, 2x GPT-5.6 Sol, 2x Fable5. The subscription fee for one week was almost completely burned out. Grok focuses on auditing, brainstorming, and overall strategy integration. Fable conducts hands-on experiments and GPT conducts in-depth analysis of EDA data. I still want to give Grok a big push, especially since I hired him to do the data analysis work after the release of 4.6. It's really super fast and rarely makes mistakes. ==Solution sharing== The main storyline ends with Imitation Learning. As a teacher who collected top teams' card decks from public game dumps, I trained a small Transformer (with two layers of d320 and encoder decoder) architecture to do three things: candidate self attention and dual seed ensemble are upgrades that have been tested and felt (receiving+3pp each in paired games), and legal mask is a basic skill to ensure a minimum. The biggest point is that it is not the model that truly determines the strength of BC, but the quality of the teacher's data. The teacher needs to pass the dual win rate barrier of global and point to point games. If a bad teacher infiltrates, they may poison the entire batch of materials; The deck only includes 52/60 or more similar families. In order to supplement state coverage, high scoring pools accept both wins and losses, while low scoring pools only accept wins to avoid losing weaker opponents. In the end, the most time was not spent on training, but on organizing data and verifying. I trained a gate matrix covering 10 opponent card groups from the teacher, played from 40 to 80 matches with Seed A/B, and weighted it according to the actual meta distribution of the gold medal area. No matter how beautiful the training indicators are, they don't count. Winning matches counts. Final hand over Dragapult+Hydreigon deck. ==Post match review== In fact, I also tried a round of RL seriously: self play, advantage weighting, actor critic, PFSP, but the results that came out later were flat or even worse. After the game, I realized the problem only after seeing the strong players' actions. In the past, it was reasonable to use pure self play in a symmetrical environment, but the cards were not; Different decks will cut the state space into pieces, and fixed decks will play themselves. Essentially, they have been practicing mirror matching, and in actual combat, most of the decks have not been seen at all. I actually did the League PK RL thing. I formed a training pool with agents from various card groups that I had trained, and covered the entire meta with training distribution. But later on, I realized that all the agents in the pool were my own, and they were fixed and not moving, so the opponent would not become stronger. After RL training, I was not learning to play cards, but learning the bot habits in the pool. The winning rate of the training pool kept increasing, but in the end, the confrontation remained stagnant. After watching the solution from the previous class after the game, I finally made up for the last piece: their coaching pool "evolves together", and experts from each deck practice and update back into the pool, which means their opponents are always newer than you think. Coaching cannot just cover meta, but also need to become stronger together with you. This is really unexpected.