deepmind/pysc2
Pysc2
pysc2 是一个 Python 接口和仿真框架，将 StarCraft II 游戏引擎连接到机器学习代理。它充当 API 包装器，将游戏内部状态公开为一组观察和动作，为研究和训练提供强化学习环境。
该框架包括用于游戏回放分析的工具，以从录制的比赛中提取数据和动作序列进行预测建模。它还提供了一个代理仿真环境，用于运行和评估单个或竞争性人工智能代理的性能。
该系统处理游戏地图配置、行为分析，并提供用于调试代理行为的手动控制接口。它将游戏状态转换为多维张量，并使用远程过程调用（RPC）框架来管理客户端与游戏引擎之间的通信。
Features
- Reinforcement Learning Environments - Provides a specialized reinforcement learning environment connecting the StarCraft II engine to machine learning agents.
- Agent Simulation Environments - Offers a system for running and testing competing AI agents to evaluate their behavior and strategic performance.
- Autonomous Agent Simulations - Implements a framework for simulating and analyzing the behavior of autonomous AI agents in a strategic environment.
- Action-Observation Cycles - Implements a synchronous loop where agents receive game state snapshots and return corresponding action commands.
- Game Automation APIs - Provides a programmatic interface for autonomous agents to receive game observations and execute actions within the engine.
- Game Simulation Environments - Offers a simulated environment for running controlled game scenarios to evaluate AI controller performance.
- Game AI Development Kits - Provides a comprehensive development kit for creating autonomous agents that operate in a real-time strategy setting.
- Model Behavioral Analysis - Facilitates the study of agent decision-making and behavioral patterns using data extracted from recorded matches.
- Replay Processing Pipelines - Parses recorded game files into sequences of observations and actions to facilitate offline supervised learning.
- Tensor Data Representations - Transforms raw game state into multi-dimensional tensors for direct consumption by deep neural networks.
- Protobuf Serialization - Uses Protocol Buffers to serialize observations and actions for consistent, strictly typed cross-language data transfer.
- Game Replay Analysis - Ships a utility for extracting observations and actions from recorded game files to review agent behavior or train models.
- Game Replay Parsers - Provides a utility for extracting data and sequences of actions from recorded game files to train predictive models.
- gRPC Client Implementations - Implements a gRPC client to manage synchronous communication of game states and actions between Python and the engine.
- Process Isolation Architectures - Runs the game engine in a separate process to ensure that engine crashes do not interrupt the agent training process.
- Reinforcement Learning - StarCraft II learning environment.
- Reinforcement Learning Environments - Environment for reinforcement learning research in StarCraft II.
- Strategy Games - Official environment for StarCraft II reinforcement learning research.
