Learning to Play Pokémon Cards with Smart AI

A team at DeNA showed how they taught a computer to win the Pokémon Trading Card Game on mobile. The game, released in late 2024, asks players to build a deck of 20 cards and battle others online. Each match ends when one side knocks out three Pokémon, but the rules are full of hidden tricks: cards can copy each other’s moves, random dice rolls decide damage, and players cannot see the opponent’s hand or deck order. Because of these layers, a simple rule‑based program could not handle every possible play. The developers turned to reinforcement learning (RL), a method where the AI learns by playing many games and adjusting its choices based on wins or losses. RL is especially useful here because new cards are added all the time; instead of rewriting code for each change, the AI can just keep learning. The system works in a loop: the game engine sends the current state to the AI, which chooses an action and returns it. The engine updates the board and feeds back new data after each turn, allowing the AI to refine its strategy. To manage the huge number of possible moves, the team broke complex actions into simple “atomic” steps—choosing to evolve a Pokémon, picking which Pokémon to evolve, and selecting the new form. An intermediary “Action Builder” stitches these steps together before sending them to the engine, keeping the AI’s decision space small.

Training required 50 million simulated matches, each lasting about a second, on six powerful servers spread across three regions. The AI used a Transformer‑style neural network with about two million parameters and no memory of past actions, keeping computation fast even for complicated states. Even with this setup, the AI sometimes played cards poorly—using a card that saves energy but then not retreating. To fix this, the engineers added penalty terms to the learning reward that target only bad actions, not the overall game state. They also increased how often useful cards appear during deck building and gave the AI extra clues (like whether energy was actually saved). After these tweaks, correct card use rose to over 96 %. Looking ahead, the team plans to explore methods that reduce wasted learning time and compare multiple game scenarios in a single state, aiming to make the AI even more efficient.

Actions