RIA: Neural Network Methods for Causal Networks

使用因果网络解决问题一直是人工智能的一个活跃研究领域，但目前的方法由于其复杂性和获得所需因果知识的难度而受到限制。本文提出，这些问题可以通过将神经网络方法直接应用于因果网络来克服。为了验证这一假设，本项目将1)推导出将概率推断表述为高度并行、严格局部计算的方法，以及2)推导出从案例数据中学习因果关系及其相关概率的方法。与过去的神经建模工作不同，这些方法将直接作用于因果网络，而不需要自己单独的网络。这些方法的性质，如系统收敛，将使用随机过程和动态系统的理论进行分析。将发展近似误差的测量。将使用不同的因果网络进行计算机模拟，以验证理论结果并选择重要参数，如退火和学习率。这项工作的主要贡献预计是开发新的，可能更有效的推理和学习方法，这些方法可以同时享受因果网络和神经网络的优势。它们可以在专家系统中用于诊断、理论形成、决策支持和其他涉及概率推理的应用。它还将提供一个使用神经网络方法直接使用人工智能表示的高级推理的具体示例。//

Use of causal networks for problem-solving has been an active research area in AI, but current methods are limited because of their complexity and the difficulty of obtaining the needed causal knowledge. It is proposed here that these problems could be overcome by applying neural network methods directly to causal networks. To test this hypothesis, this project will 1) derive methods for formulating probabilistic inferences as highly parallel, strictly local computations, and 2) derive methods for learning causal relationships and their associated probabilities from case data. Unlike past neural modeling work, these methods will directly act on causal networks without requiring their own separate networks. Properties of these methods, such as systems convergence, will be analyzed using theories of stochastic processes and dynamic systems. Measurement of approximation errors will be developed. Computer simulations using different causal networks will be performed to verify the theoretical results and to select important parameters such as annealing and learning rates. The main contribution of this work is expected to be the development of new, potentially more efficient inference and learning methods that enjoy the advantages of both causal networks and neural networks. They can be used in expert systems for diagnosis, theory formation, decision support, and other applications involving probabilistic reasoning. It will also provide a concrete example of high level inference using neural network methods that make direct use of an AI representation.//