gov - Overview on government state - allows to change government forms.game - Contains current game information - mostly irrelevant for training.dipl - Currently deprecated information on diplomacy - see player.client - Contains information on the current client communicating with the server.city - Overview on state of each individual city (improvements, production, etc.) and actions each city can take (work, unwork tiles, produce units/improvements, etc.).Both state and action_opt are themselves dictionaries describing different aspects of the game, namely: Observation is a tuple of the current state of the game and the action_options. Installation for bot developers and gym users In order to focus on a) b) and c) only, round-based games like Freeciv are a potential intermediate step for developing AI before jumping to real-time strategy games. On top of these challenges real-time strategy games due their continuous time and space add another layer of complexity (e.g., optimal formation of units, evade balistic projectiles or missiles). Hence, "naive" greedy approaches to improving a single score (e.g., Flipper, Super Mario, etc.) won't cut it. The longer an effect takes to materialize, mapping the actions/decisions to the very effect gets more complicated. needs to be more dynamic and will face more sparce option spacesĬ) More long-term consequences: In many cases, effect of actions and decisions can only be seen after several rounds, minutes of play. The key challenges when switching from board games to strategy games is:Ī) Knowledge of state of board is incomplete - therefore information has a cost: Due to fog of war (and unclear worklists, resources) position/status of the enemy and overall map is not fully known - hence risky investments (explorer teams) are required to gain informationī) Number of actors/actions is unlimited: Buildings, Cities, Units can be produced infinitely - hence the data model and learning method for neural networks etc. OpenAI and Deepmind already start tackling more complex strategy games like Dota 2 and Starcraft 2. Recent breakthroughs by AlphaZero, AlphaGo show capabilities of current systems to master "simple" board games via reinforcement learning and Montecarlo tree search (MCTS). Looking for contributors, AI and Freeciv enthusiasts. The main goal is to foster AI research using a round-based strategy game like freeciv. Freeciv-bot allows to steer freeciv via Python and build automated bots that can ultimately beat human players.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |