Autonomous Repairing Ability in Computers

随着未来计算机硬件规模的扩大，对硬件故障的高鲁棒性的需求大大增加。本研究的目的是提出对故障具有高鲁棒性的计算机，发生在他们的运行时间以及在制造过程中。在本研究中，神经计算机被选为目标计算机。神经计算机具有很高的容错性，因为它模仿了生物神经系统。然而，关于它的容错性的研究还不多见。特别是SOM（Self-Organizing Map）作为一种很有前途的神经计算机体系结构，其容错性还没有得到充分的研究，本文从理论和实验两方面，利用真实的神经计算机对SOM的自主修复能力进行了评估。研究结果表明：1）SOM神经元的故障会产生一个很长的伪稳定状态，但伪稳定状态之后最终会达到全局有序状态。提出了一种缩短伪稳态周期的方法，并对其有效性进行了定量分析。2）从缺陷SOM模型出发，推导出缺陷神经元的临界卡滞输出rho_c。如果缺陷神经元的输出rho_d远小于rho_c，则很难达到全局有序状态。然而，在现有的神经计算机中，这种不希望的情况发生的概率约为所有卡滞输出的6%。3）对有缺陷的SOM进行了实验，实验结果与上述理论预测一致，表明理论评估是正确的。所得结果和准则将用于未来的神经计算机和容错神经计算机的设计。

Need for high robustness against hardware faults increases considerably in the future computers as the hardware sizes of them enlarge. The goal of this research is to propose the computers which have high robustness against faults happen in their run-time as well as in the fabrication process. In this research, neuro-computer is chosen as a target computer. High fault tolerance is expected in the neuro-computer because it mimics the biological neural system. However, not so many results have been repoted concerning its fault tolerance. Especially, the fault tolerance of SOM (Self-Organizing Map) which is one of the promising neuro-computer architectures has not been explored yet.The autonomous repairing ability of the SOM has been evaluated theoretically and experimentally, using the real neuro-computer. The research conclusions are :1) an extremely long pseudo stable state emerges because of the faulty neurons in the SOM.However, the global ordering state is achieved eventually after the pseudo stable state. A technique to shorten the pseudo stable state period has been proposed and its effectiveness has been analyzed quantitatively.2) the critical-stuck-output rho_c of the defective neuron has been derived from the defect SOM model. If the defective neuron's output rho_d is much smaller than rho_c, the global-ordering state can hardly be achieved. However, the probability that this undesired case happens is about 6% of all stuck-outputs in the present neuro-computer.3) Experiments on the defective SOM have been carried out, and the results have been well agreed with the above theoretical predictions, indicating that the theoretical evaluation is correct. The obtained results and criteria will be utilized for the future neuro-computers and fault-tolerant neuro-computers designs.