智能自适应系统：最佳实践与选择

需积分: 5 173 浏览量更新于2024-07-17 收藏 5.1MB PDF 举报

"Do Smart Adaptive Systems Exist? Best Practice for Selection and Combination of Intelligent Methods" 是一本探讨智能自适应系统存在的可能性以及如何选择和组合智能方法的书籍。这本书属于模糊性和软计算研究系列，由 Bogdan Gabrys、Kauko Leiviskä 和 Jens Strackeljan 编辑，Janusz Kacprzyk 担任主编。智能自适应系统是当前信息技术领域的一个重要概念，这类系统能够根据环境变化和用户需求自动调整其行为和性能。它们通常包含一系列先进的算法和技术，如人工智能、机器学习、模糊逻辑、进化计算等，旨在实现高效、灵活且自主的决策和操作。书中的讨论可能涵盖了以下几个关键知识点： 1. **智能自适应系统的定义**：书中首先定义了智能自适应系统，这包括了系统应具备的能力，比如感知环境变化、学习和自我优化、自主决策等。 2. **智能方法的选择与组合**：为了构建有效的智能自适应系统，选择适合的智能算法至关重要。书中可能会详细讨论如何评估各种智能方法（如神经网络、遗传算法、模糊逻辑系统等）的优点和局限性，并提供最佳实践指导，以实现不同技术的最优组合。 3. **应用案例分析**：通过对具体应用案例的深入分析，作者可能探讨了这些系统是否满足智能自适应的要求，以验证这些技术在实际场景中的效果。 4. **模糊逻辑与软计算**：作为该系列的一部分，书中的内容可能涉及模糊逻辑和软计算的基本原理，以及它们在构建智能系统中的作用，特别是处理不确定性和不精确信息的能力。 5. **进化计算与数据挖掘**：进化计算（如遗传算法、粒子群优化等）是智能系统中用于解决复杂优化问题的工具，而数据挖掘则为系统提供了从大量数据中发现模式和知识的能力。这些可能也是书中讨论的焦点。 6. **多智能体系统设计**：书中可能还包括了关于如何设计和实施智能多代理系统的内容，强调了协作、通信和分布式决策在复杂系统中的重要性。 7. **成本效益分析**：在实际应用中，智能系统的效益必须与投入的成本进行比较。书中的某些章节可能涉及成本效益分析和模糊决策理论，帮助读者理解如何在模糊环境中做出合理的决策。 8. **系统研发与实施**：除了理论探讨，书籍也可能涵盖了系统开发过程中的实践经验，包括设计、测试、调试和维护等方面的策略和技巧。通过深入阅读本书，读者可以全面了解智能自适应系统的设计理念、技术选择和实际应用，对于从事相关领域研究或开发的人员来说，是一份宝贵的参考资料。

1 Do Smart Adaptive Systems Exist? – Introduction 7

of some features of the environment. Examples are systems that must be

ported from one plant to another without a need to explicitly perform a

porting of its main parameters, or a ﬁnancial application that must be

ported from a speciﬁc market to a similar one (e.g. a diﬀerent geographical

location).

3. Adaptation to new/unknown application. This level is the most futuris-

tic one, but its open problems have been addressed already by a number

of researchers, especially in the Machine Learning and Hybrid Intelligent

Systems ﬁelds where, starting from very limited information about the

problem it is possible to build a system through incremental learning.

The working deﬁnition adopted within EUNITE is that “smart” implies

that intelligent techniques must be involved in the adaptation of a system.

That is, to be a smart adaptive system, a system must be adaptive to one

of the levels deﬁned above and must utilise one or more of the intelligent

techniques mentioned in the previous section. This, though very practical, is

somewhat a very narrow deﬁnition of the word “smart” and therefore systems

conforming to this deﬁnition are denoted as narrowly smart.

On the other hand, on a more general level, the word “smart” can be taken

to have a meaning closer to a traditional dictionary deﬁnition of “intelligent”:

i.e. the ability to learn or understand or to deal with new or trying situations;

or the ability to apply knowledge to manipulate one’s environment or to think

abstractly as measured by objective criteria (as tests) [The Merriam-Webster’s

Dictionary]. This is beyond even level 3 adaptation as deﬁned above and we

denote it as generally smart.

While examples of levels 1 and 2 applications can be found in abundance

and many of such applications are discussed in the second part of the book

(Chaps. 10 to 17) the level 3 problems and issues are much more challenging.

So do any attempts have been made to construct such systems? Have any

general requirements/desirable characteristics for such systems been speciﬁed?

What are the challenges and problems?

Let us now focus on an example of one such general framework and il-

lustrate some of the main issues and attempt to answer some of the above

questions.

1.4 General Framework for SAS

While searching for answers to the question whether smart adaptive systems

exist we have come across the following deﬁnition. Paraphrasing Kasabov

who was describing knowledge engineering on-line and emerging intelligence,

we could also say that a smart adaptive system is a system “that develops its

structure and functionality in a continuous, self-organised, adaptive, interac-

tive way. It can use many sources of incoming information and can perform

intelligent tasks, such as adaptive pattern recognition, concept formation, lan-

guage learning, reasoning with uncertainty, decision making, and many more”.

8 B. Gabrys

And in fact Kasabov [27] listed seven major requirements of future intelli-

gent systems and proposed an open framework called Evolving Connectionist

System (ECOS) [27, 28], some advanced aspects of which are discussed in

Chap. 9. We will now use both the requirements and the ECOS framework as

the basis for our discussion as, though with a number of open problems still

to be solved, they are the closest yet to the level 3 smart adaptive system as

deﬁned in the previous section.

Seven major requirements for future intelligent/smart adaptive systems:

1. SAS should have an open, extendable and adjustable structure. This means

that the system has to have the ability to create new inputs and outputs,

connections, modules etc. while in operation. It should also be able to

accommodate in an incremental way all the data that is know and will be

known in the future about the problem.

2. SAS should adapt in an on-line, incremental, life-long mode where new

data is used as soon as it is available.

3. SAS should be able to learn fast from large amount of data ideally in an

“one-pass” training mode.

4. SAS should be memory-based with an ability to add, retrieve and delete

individual pieces of data and information.

5. SAS should be able to improve its performance via active interaction with

other systems and with the environment in a multi-modular, hierarchical

fashion.

6. SAS should represent adequately space and time at their diﬀerent scales,

inner spatial representation, short- and long-term memory, age, forgetting,

etc.

7. SAS should be able to self-improve, analyse its own performance and ex-

plain what it has learnt about the problem it is solving.

While many of the above issues have been acknowledged and addressed by

researchers working with intelligent technologies since their conceptions, the

focus tends to be on the individual or small subsets of the above listed points.

The reason for this is that any of the issues listed are often diﬃcult enough in

themselves and either the structure, learning or the knowledge representation

of the existing techniques are not adequate or ﬂexible enough to cover all of

them. And as correctly pointed out in [27] it is not likely that a truly smart

adaptive system can be constructed if all of the above seven requirements are

not met and therefore radically new methods and systems are required.

One such attempt at proposing a model or a general framework that ad-

dresses, at least in principle, all seven requirements is the ECOS framework.

ECOS are multi-level, multi-modular, open structures which evolve in time

through interaction with environment.

There are ﬁve main parts of ECOS which include:

1. Presentation part. In this part an input ﬁltering, preprocessing, feature se-

lection, input formation etc. are performed. What is interesting and some-

thing that would naturally be a consequent of acquiring new data and

1 Do Smart Adaptive Systems Exist? – Introduction 9

information is that the dimensionality of inputs can be variable. This usu-

ally can pose serious problems with many existing statistical, neural or

machine learning techniques.

2. Representation and memory part. In the original ECOS this part is formed

by a multi-modular, evolving structure of neural network modules used

for storing input patterns (information). In fact this part could consist of

various functional modules from various domains as long as they could be

created and optimised in an on-line manner (i.e. decision trees, neuro-fuzzy

modules etc.).

3. Higher level decision part. This part consisting of several modules is quite

critical as it assumes that a suitable feedback from environment is received

and decisions about the functioning and adaptation of the whole ECOS

are made here. Many unresolved research issues and important questions

are still to be tackled in this part and will pose signiﬁcant challenge in the

near future.

4. Action part. In this part the decisions from the decision module are passed

to the environment.

5. Self-analysis and rule extraction modules. This part is used for the extrac-

tion of the information from the representation and decision modules and

converts it into diﬀerent forms of rules, associations etc. which may be use-

ful in the process of interpretability and transparency of the functionality

and knowledge stored in ECOS model.

As mentioned in Chap. 9 there are a number of realisations of ECOS

framework especially evolving, hierarchically structured neuro-fuzzy systems

with the ability to grow and shrink as required. An example of such a system

also include the general fuzzy min-max neural networks (GFMM) [18, 19, 20,

21, 22, 23] in which a number of requirements for SAS have been addressed

and are being further developed and investigated.

An example of the process model for developing adaptive systems using

machine and reinforcement learning with some components which are similar

to the above ECOS framework are discussed in Chap. 5 while another example

of typical general steps or functional elements required for SAS application

are highlighted in Chap. 2.

So does this mean that the problem of smart adaptive systems has been

solved? The answer is NO. Though signiﬁcant advances have been made there

are still important issues and challenges requiring intensive research.

Let us now try a take a closer look at some of those critical issues and

challenges within the context of the seven requirements and the listed ECOS

framework components.

10 B. Gabrys

1.5 Achievements, Critical Issues and Challenges

1.5.1 Data Preparation and Preprocessing

One of the most underrated but absolutely critical phases of designing SAS

are data preparation and preprocessing procedures [43]. In order to highlight

the importance of this stage, as part of the introductory material to the main

two parts of the book, Chap. 3 has been entirely dedicated to these issues.

Why are they so important? Since the changing environments requiring

adaptation usually manifest themselves in changes in the number and type of

observed variables (see requirement 1 and presentation part of ECOS), it is

particularly important/critical to know the implications it could have on pre-

processing (e.g. dealing with noise, missing values, scaling, normalisation etc.)

and data transformation techniques (e.g. dimensionality reduction, feature

selection, non-linear transformations etc.) which in turn have an enormous

impact on the selection and usage of appropriate modelling techniques. It is

even more critical as this stage in a SAS application should be performed in

an automatic fashion while currently in practise a vast majority of the data

cleaning, preprocessing and transformation is carried out oﬀ-line and usually

involves a considerable human expertise.

1.5.2 Adaptation and Learning

Not surprisingly the adaptation and learning algorithms should play a promi-

nent role in SAS and in fact in one form or another are represented in all

seven requirements.

There are currently two main groups of search, adaptation and learning

algorithms: local (often gradient based) search/learning algorithms and global

(stochastic) approaches.

The main learning algorithms in the ﬁrst group have been developed over

the years in statistics, neural networks and machine learning ﬁelds [4, 5, 13, 44,

45, 52]. Many of the methods are based on the optimization/minimization of a

suitable cost, objective or error function. Often those algorithms also require

the cost function to be diﬀerentiable and many gradient based optimization

techniques have been used to derive learning rules on such basis. Unfortunately

such diﬀerentiable cost functions do not always exist.

Evolutionary algorithms and other stochastic search techniques [16, 24,

26, 30] provide a complementary powerful means of a global search often

used where gradient based techniques fail. As mentioned earlier these global

searches usually are less computationally eﬀective then the gradient based

learning techniques.

Due to their complementary nature it is more and more common that both

types of algorithms are used in combination for SAS as illustrated in Chaps. 8

and 9.

1 Do Smart Adaptive Systems Exist? – Introduction 11

Another distinction between diﬀerent learning algorithms which all are

needed in SAS general framework is based on the availability and type of the

training feedback signal and fall into three broad categories:

• Supervised learning [5, 45] – also known as learning with a teacher where

both the input patterns and the desired target signal are available.

• Unsupervised learning [5, 45] – which does not assume the presence of the

feedback signal and attempts to group the input patterns on the basis of

a suitably chosen similarity measure.

• Reinforcement learning [49] – in which the external feedback signal is

restricted to a discrete, delayed reward and punishment signals and the

correct output is not known.

While a lot of research has been carried out in each of the above mentioned

groups and further details and examples of some representative algorithms can

be found in Chaps. 4 and 5, the nature of real problems requiring adaptation

makes it also necessary that these diﬀerent types of algorithms be combined

[17, 18, 22, 42]. The combination of these learning algorithms is a vigorously

pursued research area on its own and they are particularly relevant to the

SASrequirements1,2,3,5and6.

With respect to the general issue of adaptive learning there are two perti-

nent problems which have received a lot of attention in the literature:

• Stability/plasticity dilemma which is associated with on-line, incremental

model parameter learning and adaptation. In simple terms the problem

relates to the ability of SAS to follow the changes and react quickly to

such changes versus ability to ﬁlter out noisy data and exhibit stable,

good overall performance.

• Bias/variance dilemma which is associated with model structure learning

and ﬁnding and maintaining the right model complexity. In other words if

the model is too simple it will not be able to capture the complex nature

of the problem (often resulting in stable but over-generalising models with

high bias) while too complex model could result in data overﬁtting and

poor generalisation characteristics (models generated are often too ﬂexible

which manifests itself in a high variance error component).

With regard to adaptive parameter learning and stability/plasticity

dilemma one can distinguish two main learning modes present in ECOS/SAS

frameworks:

• On-line/incremental learning - which refers to the ability to learn new

data without destroying old data and learning while system is in oper-

ation (dominated by various supervised, unsupervised and reinforcement

techniques from ANN and ML [5, 13, 28, 45, 44]).

• Life long learning - which refers to the ability of the system to learn during

its entire existence. Growing and pruning (techniques from ML, ANN and

FS [18, 22, 28, 44, 52]) are the primary mechanisms in this type of learning.

剩余365页未读，继续阅读

喜欢轻松

粉丝: 0
资源: 8

智能自适应系统：最佳实践与选择

Neurofuzzy-Adaptive-Learning-.rar_单片机开发_PDF_

Backstepping-Adaptive-Fuzzy-Control-for-two-link-_matlab例程_PDF_

ExtremeLearningMachine资源共享-Research-on-computational-intelligence-algorithms-with-adaptive-_2013_Neuroc.pdf

single-neuron-adaptive-PID.zip_PID-NTC_adaptive control_adaptive

Functional+Safety+in+Microprocessors+with+MICROSAR+Adaptive.pdf

An_Adaptive_Distribution-Free_Algorithm_for_the_Ne.pdf

lms_adaptive-_filtering.rar_LMS自适应

adaptive-routing-proto.tar.gz_行业发展研究_PDF_

adaptive_blind.zip_Blind OFDM_adaptive-blind_blind_ofdm_ofdm bli

Adaptive-background-mixture-models-for-real-time-_For Real_adapt

最新资源