【计算机科学】【2012.07】研究人类行为的人工神经网络

本文为美国爱荷华大学（作者：Mohammad Hindi Bataineh）的硕士论文，共179页。

工业和学术界绝大多数产品和工序都需要人与人之间的互动。因此，数字人体模型（DHMs）对于改进设计、预防伤害以及更好地理解人类行为变得至关重要。尽管DHM领域的许多技术正在不断成熟，但仍有非常多的改进机会，特别是在姿态和运动预测方面。因此，本论文探讨使用人工神经网络（ANN）来改善预测能力，以及更好地理解人类如何及为何会产生这样的行为。

关于运动预测，最具挑战性的机会之一是涉及到计算速度。特别地，当考虑动态运动预测时，潜在的优化问题可能很庞大，并且计算复杂。尽管目前用于预测人体姿态的基于优化的工具相对快速和准确，因此不需要太多的改进，但是姿态预测通常比运动预测更容易处理，并且能够提供测试的基础，以揭示运动预测中潜在的问题。因此，本文将ANN应用于姿态预测，以发现潜在的问题。此外，直接使用人工神经网络进行姿态预测，提供了使用人工神经网络来预测行为度量（PMs）的最适当组合的初始步骤——驱使人类行为的原始动力。PM是姿态预测问题中最小化的代价函数，通常根据任务需求进行手动选择。这可能是使用姿态预测时最显著的障碍。用户如何知道应使用哪一种PM？神经网络为解决这个问题提供了工具。

本文假设人工神经网络可以通过训练进行快速准确地预测人体运动、预测人体姿态（同时考虑外力的影响），并确定用于姿态预测的PMs的最恰当组合。这样的能力将反过来为研究人类行为提供一种新工具。通过初步实验，发现广义回归神经网络（GRNN）是DHM应用中最有效的人工神经网络类型。设计了一种半自动化的方法来简化神经网络构建、训练和测试过程以及网络参数，从而促进了DHM的具体应用。

在运动预测方面，人工神经网络的应用是成功的。研究结果表明，在不降低计算精度的情况下，计算时间从1 - 40分钟减少到几分之一秒。关于姿态预测，ANN再次被发现是有效的。然而，某些运动预测任务的潜在问题也被发现，并阐明了人工神经网络未来发展的必要性。最后，使用GRNN开发了一种决策引擎，用于自动选择四种人类PMs，并被证明是非常有效的。为了训练这种新方法，使用一种新的优化公式从已有的运动捕获数据中提取PM权重。最终，这项工作将在一般的虚拟环境中自动地、逼真地驱动预测出DHMs。

The vast majority of products and processes in industry and academiarequire human interaction. Thus, digital human models (DHMs) are becomingcritical for improved designs, injury prevention, and a better understanding ofhuman behavior. Although many capabilities in the DHM field continue to mature,there are still many opportunities for improvement, especially with respect toposture- and motion-prediction. Thus, this thesis investigates the use ofartificial neural network (ANN) for improving predictive capabilities and forbetter understanding how and why human behave the way they do. With respect tomotion prediction, one of the most challenging opportunities for improvementconcerns computation speed. Especially, when considering dynamic motionprediction, the underlying optimization problems can be large and computationallycomplex. Even though the current optimization-based tools for predicting humanposture are relatively fast and accurate and thus do not require as much improvement,posture prediction in general is a more tractable problem than motion predictionand can provide a test bead that can shed light on potential issues with motionprediction. Thus, I investigate the use of ANN with posture prediction in orderto discover potential issues. In addition, directly using ANN with postureprediction provides a preliminary step towards using ANN to predict the mostappropriate combination of performance measures (PMs) - what drives humanbehavior. The PMs, which are the cost functions that are minimized in theposture prediction problem, are typically selected manually depending on thetask. This is perhaps the most significant impediment when using postureprediction. How does the user know which PMs should be used? Neural networksprovide tools for solving this problem.

Thisthesis hypothesizes that the ANN can be trained to predict human motion quicklyand accurately, to predict human posture (while considering external forces),and to determine the most appropriate combination of PM(s) for postureprediction. Such capabilities will in turn provide a new tool for studyinghuman behavior. Based on initial experimentation, the general regression neuralnetwork (GRNN) was found to be the most effective type of ANN for DHMapplications. A semi-automated methodology was developed to ease networkconstruction, training and testing processes, and network parameters. This inturn facilitates use with DHM applications.
Withregards to motion prediction, use of ANN was successful. The results showedthat the calculation time was reduced from 1 to 40 minutes, to a fraction of a secondwithout reducing accuracy. With regards to posture prediction, ANN was again foundto be effective. However, potential issues with certain motion-prediction taskswere discovered and shed light on necessary future development with ANNs.Finally, a decision engine was developed using GRNN for automatically selectingfour human PMs, and was shown to be very effective. In order to train this newapproach, a novel optimization formulation was used to extract PM weights frompre-existing motion capture data. Eventually, this work will lead toautomatically and realistically driving predictive DHMs in a general virtualenvironment.

1 引言

2 人工神经网络：项目背景

3 采用ANN进行人类建模的新方法

4 基于神经网络的运动预测

5 基于神经网络的姿态预测

6 人类行为度量的决策引擎设计

7 总结与结论

附录A 人类模型

附录B 所有任务的训练示例列表

下载英文原文地址：

http://page2.dfpan.com/fs/5lcja22142910658255/

更多精彩文章请关注微信号：