SafeLife

该数据集包含了为训练智能体在康威生命游戏中执行任务时尽量减少副作用而生成的一系列环境。此外，该数据集被用于评估五种不同条件（PPO、DQN、AUP、AUP投影、朴素方法）在“追加生成”和“修剪静态简单”两项任务上的表现。规模上，该数据集包含多个环境，每个任务都有8个环境作为课程设置。总体任务目标是训练强化学习智能体完成任务的同时避免产生副作用。

This dataset consists of a collection of environments created to train AI Agents to minimize side effects when executing tasks within Conway's Game of Life. In addition, this dataset is utilized to evaluate the performance of five distinct training paradigms (PPO, DQN, AUP, AUP Projection, naive method) across two tasks: "Append Generation" and "Prune Static Simple". In terms of scale, the dataset includes multiple environments, with 8 environments serving as the curriculum for each individual task. The overarching goal of the tasks is to train reinforcement learning agents to complete their assigned tasks while avoiding unintended side effects.