Deep Variational Autoencoders for NPC Behaviour Classification
Deep Variational Autoencoders for NPC Behaviour Classification
复制标题
用于 NPC 行为分类的深度变分自动编码器
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
V. Bulitko
中科院分区：
文献类型：
--
作者：
E. S. Soares;V. Bulitko
程序内容生成（PCG）可以在视频游戏中创建新颖的、玩家特定的内容，包括AI控制的不可玩角色（NPC）的行为。在这里，我们提出了我们的第一个结果，比较无监督和有监督的机器学习程序生成的NPC行为。使用人工生命环境作为视频游戏的替身，我们运行人工进化并生成具有各种行为的AI代理。然后，我们在常见的进化行为上训练深度变分自编码器，并测量其在检测训练过程中看不见的行为方面的功效。作为参考，我们使用一个以监督方式训练的现成的深度网络来检测训练过程中看到和看不到的行为。初步结果表明，即使训练集包含几种类型的行为的混合物而没有适当的标签，也具有良好的性能。
Procedural content generation (PCG) can create novel, player-specific content in video games, including behaviours of AI-controlled non-playable characters (NPC). Here we present our first results on comparing unsupervised and supervised machine learning for procedurally generated NPC behaviours. Using an artificial life environment as a stand-in for a video game, we run artificial evolution and generate AI agents with various behaviours. We then train deep variational autoencoders on commonly evolved behaviour and measure its efficacy in detecting behaviours unseen during training. As a reference, we use an off-the-shelf deep network trained in a supervised manner to detect behaviours both seen and unseen during its training. Preliminary results demonstrate promising performance that holds even when the training set contains a mixture of several types of behaviours without proper labels.