Skip to content

Bonus Unit 3 — Agents in Games (Pokémon)

Overview

This bonus unit applies agent techniques to a game environment — specifically a Pokémon battle simulator. It is a fun way to test agent decision-making in a fully observable, turn-based setting.

Why games?

Games provide:

  • Clear win/lose signal — easy to evaluate agent performance
  • Structured action space — finite set of valid moves
  • Reproducible environments — easier to debug
  • No real-world consequences — safe to experiment

The Pokémon environment

The course uses Pokémon Showdown via the poke-env library:

pip install poke-env

from poke_env.player import Player

class MyAgent(Player):
    def choose_move(self, battle):
        # Use the LLM to decide which move to make
        context = self._build_context(battle)
        move_name = self._llm_decide(context)
        return self.create_order(battle.available_moves[0])  # simplified

Agent design considerations

  • State representation — how to describe the battle state to the LLM
  • Action grounding — ensuring the LLM picks a valid move
  • Memory — tracking HP, status conditions, and previous turns
  • Strategy — type matchups, priority moves, switching

Notes & experiments

Add your game agent notes, battle logs and win-rate experiments here.