基于生物进化的可恢复回声状态网络模型

Restorable Echo State Network Based on Biological Evolution

为解决回声状态网络储备池在遭受随机故障和蓄意攻击等复杂情况下的适应性问题,提出了一种具有生物进化特征的可恢复回声状态网络—3DP-RESN.基于优先匹配的复制、新增加连接的变异和新增加连接的死亡进化策略,3DP-RESN能够实现从被破坏的网络拓扑中自恢复.将3DP-RESN、传统ESN(CESN)和被破坏的ESN(DESN)应用于NARMA系统、Henon映射和figure8这3种非线性时间序列逼近任务.实验结果表明,当储备池发生故障时,3DP-RESN对于3种时间序列的预测精度明显优于DESN,接近甚至高于未遭受储备池故障的CESN,尤其在figure8实验中,3DP-RESN与CESN、DESN相比,预测精度分别提高了30.56%和7.01%.此外,3DP-RESN的短期记忆能力也接近于CESN,因此,3DP-RESN具有强大的自适应恢复能力.

To solve adaptability problems of the reservoirs of echo state network in complicated conditions,such as suffering from random faults and deliberate attacks,a restorable echo state network with biological evolution characteristics—3DP-RESN was proposed.The 3DP-RESN was designed to be able to recover automatically from destroyed network topology based on the evolution strategies of preferentially matched duplication,newly added connection-oriented divergence and newly added connection-oriented death.In experiments,3DP-RESN,classic ESN（CESN）and destroyed ESN（DESN）are applied to approximating three kinds of nonlinear time series,i.e.,the NARMA system,Henon map and figure8.Experimental results show that,when reservoirs suffer from failure,for three kinds of time series,the prediction accuracy of 3DPRESN significantly outperforms DESN,and is close to or even higher than that of CESN which has not suffered from failure.Especially in the experiment of figure8,compared with CESN and DESN,the prediction accuracy of 3DP-RESN is improved by 30.56% and 7.01% respectively.Besides,the short-term memory capacity of the 3DP-RESN is also close to that of CESN.Hence,3DP-RESN can possess strongly adaptive self-recovery capacity.