sgai-team-alien
收藏资源简介:
Team Alien Deluxe RL数据集是一个用于强化学习数据收集和分析的数据集,专门为Team Alien的Deluxe RL项目设计。数据集主要包含基础游戏模式下的实验数据,分为两个基线代理(harish和karunya)的结果。数据集由多个文件组成:episodes.csv文件包含约1000个完整游戏回合的统计信息,记录每个回合的模型标识符、运行ID、随机种子、结束原因、各种动作数量(保存、跳过、压碎、扰乱、弹出)、回合数、保存和杀死数量、保存值和奖励等。steps.csv文件包含约40000个步骤级别的详细信息,记录每个步骤的模型标识符、运行ID、随机种子、步骤索引、回合索引、决策类型、执行动作、选择的动作索引、Q值、动作掩码、贪婪标志、观察状态、人形特征、真实状态、载具状态、载具概率、显示载具百分比、剩余时间、使用槽位、总槽位、乘客数量、健康救护车、受伤救护车、僵尸救护车、尸体救护车、健康载具救护车、受伤载具救护车、救护车价值、载具后验概率、安全扰乱概率、弹出槽位、弹出真实状态、弹出载具状态、弹出载具概率、弹出显示百分比等。此外还包括评估摘要文件(eval.md)和训练好的模型检查点文件(harish.pth和karunya.pth)。Deluxe游戏模式目录当前为空,预留用于未来扩展。该数据集适用于强化学习算法分析、游戏AI行为研究、决策过程建模等任务。
The Team Alien Deluxe RL dataset is a dataset designed for reinforcement learning data collection and analysis, specifically for Team Aliens Deluxe RL project. It primarily contains experimental data from the basic game mode, divided into results from two baseline agents (harish and karunya). The dataset consists of multiple files: episodes.csv includes statistics for approximately 1000 complete game episodes, recording each episodes model identifier, run ID, random seed, termination reason, counts of various actions (save, skip, crush, disrupt, pop), episode number, save and kill counts, save value, and reward. steps.csv contains detailed step-level information for about 40,000 steps, recording each steps model identifier, run ID, random seed, step index, episode index, decision type, executed action, selected action index, Q-value, action mask, greedy flag, observation state, humanoid features, true state, vehicle state, vehicle probability, displayed vehicle percentage, remaining time, used slots, total slots, passenger count, healthy ambulance, injured ambulance, zombie ambulance, corpse ambulance, healthy vehicle ambulance, injured vehicle ambulance, ambulance value, vehicle posterior probability, safe disruption probability, pop slots, pop true state, pop vehicle state, pop vehicle probability, pop displayed percentage, etc. Additionally, it includes an evaluation summary file (eval.md) and trained model checkpoint files (harish.pth and karunya.pth). The Deluxe game mode directory is currently empty, reserved for future expansion. This dataset is suitable for tasks such as reinforcement learning algorithm analysis, game AI behavior research, and decision process modeling.
数据集概述
数据集名称: Team Alien Deluxe RL Dataset
许可证: MIT License
数据集地址: https://huggingface.co/datasets/maxzhangalt/sgai-team-alien
该数据集用于 Team Alien 的 Deluxe RL 数据收集与分析,主要包含强化学习游戏模式下的训练数据、评估结果和模型权重。
仓库结构
├── base/ │ ├── baselineRL_harish/ │ │ ├── episodes.csv │ │ ├── steps.csv │ │ └── eval.md │ ├── baselineRL_karunya/ │ │ ├── episodes.csv │ │ ├── steps.csv │ │ └── eval.md │ └── models/ │ ├── harish.pth │ └── karunya.pth └── deluxe/
基础游戏模式(Base Game Mode)
base/ 目录包含基础游戏模式下的数据集、评估结果和训练好的模型。
基线目录
base/baselineRL_harish/base/baselineRL_karunya/
每个基线目录包含以下三个文件。
episodes.csv - 回合级统计数据
包含每局完整游戏的回合级统计量。
- 行数: 约 1,000 行
- 回合标识符:
run_id - 列字段:
model_id, run_id, seed, end_reason, n_save, n_skip, n_squish, n_scram, n_eject, n_turns, saved, killed, saved_value, reward
每一行代表一局完整的游戏回合。
steps.csv - 步骤级详细信息
包含每个回合中每一步的行动和决策详细数据。
- 行数: 约 40,000 行
- 回合标识符:
run_id - 步骤标识符:
step_index - 列字段:
model_id, run_id, seed, step_index, turn_index, decision_kind, action, chosen_action_index, q_values, action_mask, greedy, obs_json, humanoid_fp, true_state, is_carrier, carrier_prob, displayed_carrier_pct, remaining_time, slots_used, slots_total, n_passengers, amb_healthy, amb_injured, amb_zombie, amb_corpse, amb_carrier_healthy, amb_carrier_injured, ambulance_value, carrier_posterior, p_safe_scram, ejected_slot, ejected_true_state, ejected_is_carrier, ejected_carrier_prob, ejected_displayed_pct
可通过 run_id 分组,重建并分析每个回合内完整的决策序列。
eval.md - 评估摘要
包含评估总结和终端指标,用于快速查看每个模型的整体行为和性能。
训练好的模型
base/models/ 目录包含训练好的强化学习模型检查点:
base/models/ ├── harish.pth └── karunya.pth
豪华游戏模式(Deluxe Game Mode)
deluxe/ 目录预留用于存放豪华游戏模式下生成的数据集、评估结果和训练模型。
当前状态: 该目录为空。待添加豪华模式数据时将更新数据集说明。




