1 Introduction to evolutionary computation
1.1 Introductory remarks
As a recognized field, evolutionary computation is quite young. The term itself was invented as recently as 1991, and it represents an effort to bring together researchers who have been following different approaches to simulating various aspects of evolution. These techniques of genetic algorithms (Chapter 7), evolution strategies (Chapter 8), and evolutionary programming (Chapter 9) have one fundamental commonality: they each involve the reproduction, random variation, competition, and selection of contending individuals in a population. These form the essential essence of evolution, and once these four processes are in place, whether in nature or in a computer, evolution is the inevitable outcome (Atmar 1994). The impetus to simulate evolution on a computer comes from at least four directions.
繁殖、变异、竞争、选择是四大共通点。
1.2 Optimization
Evolution is an optimization process (Mayr 1988, p 104). Darwin (1859, ch 6) was struck with the ‘organs of extreme perfection’ that have been evolved, one such example being the image-forming eye (Atmar 1976). Optimization does not imply perfection, yet evolution can discover highly precise functional solutions to particular problems posed by an organism’s environment, and even though the mechanisms that are evolved are often overly elaborate from an engineering perspective, function is the sole quality that is exposed to natural selection, and functionality is what is optimized by iterative selection and mutation.
？？？？？
It is quite natural, therefore, to seek to describe evolution in terms of an algorithm that can be used to solve difficult engineering optimization problems. The classic techniques of gradient descent, deterministic hill climbing, and purely random search (with no heredity) have been generally unsatisfactory when applied to nonlinear optimization problems, especially those with stochastic, temporal, or chaotic components. But these are the problems that nature has seemingly solved so very well. Evolution provides inspiration for computing the solutions to problems that have previously appeared intractable. This was a key foundation for the efforts in evolution strategies (Rechenberg 1965, 1994, Schwefel 1965, 1995).
随机和梯度下降在求解非线性问题中都不好，自然看起来解决的好。
1.3 Robust adaptation
The real world is never static, and the problems of temporal optimization are some of the most challenging. They require changing behavioral strategies in light of the most recent feedback concerning the success or failure of the current strategy. Holland (1975), under the framework of genetic algorithms (formerly called reproductive plans), described a procedure that can evolve strategies, either in the form of coded strings or as explicit behavioral rule bases called classifier systems (Chapter 12), by exploiting the potential to recombine successful pieces of competing strategies, bootstrapping the knowledge gained by independent individuals. The result is a robust procedure that has the potential to adjust performance based on feedback from the environment.
robust是能根据反馈调节策略的能力。？？？？？
1.4 Machine intelligence
Intelligence may be defined as the capability of a system to adapt its behavior to meet desired goals in a range of environments (Fogel 1995, p xiii). Intelligent behavior then requires prediction, for adaptation to future circumstances requires predicting those circumstances and taking appropriate action. Evolution has created creatures of increasing intelligence over time. Rather than seek to generate machine intelligence by replicating humans, either in the rules they may follow or in their neural connections, an alternative approach to generating machine intelligence is to simulate evolution on a class of predictive algorithms. This was the foundation for the evolutionary programming research of Fogel (1962, Fogel et al 1966).
智力是是一个系统为适应环境变化做出的适应性行为，需要预测采取合适举措。进化创造的个体智力越来越高。？？？？
1.5 Biology
Rather than attempt to use evolution as a tool to solve a particular engineering problem, there is a desire to capture the essence of evolution in a computer simulation and use the simulation to gain new insight into the physics of natural evolutionary processes (Ray 1991) (see also Chapter 4). Success raises the possibility of studying alternative biological systems that are merely plausible images of what life might be like in some way. It also raises the question of what properties such imagined systems might have in common with life as evolved on Earth (Langton 1987). Although every model is incomplete, and assessing what life might be like in other instantiations lies in the realm of pure speculation, computer simulations under the rubric of artificial life have generated some patterns that appear to correspond with naturally occurring phenomena.
通过计算机模拟返回去对自然进化加深理解。
1.6 Discussion
The ultimate answer to the question ‘why simulate evolution?’ lies in the lack of good alternatives. We cannot easily germinate another planet, wait several millions of years, and assess how life might develop elsewhere. We cannot easily use classic optimization methods to find global minima in functions when they are surrounded by local minima. We find that expert systems and other attempts to mimic human intelligence are often brittle: they are not robust to changes in the domain of application and are incapable of correctly predicting future circumstances so as to take appropriate action. In contrast, by successfully exploiting the use of randomness, or in other words the useful use of uncertainty, ‘all possible pathways are open’ for evolutionary computation (Hofstadter 1995, p 115). Our challenge is, at least in some important respects, to not allow our own biases to constrain the potential for evolutionary computation to discover new solutions to new problems in fascinating and unpredictable ways. However, as always, the ultimate advancement of the field will come from the careful abstraction and interpretation of the natural processes that inspire it.
为什么要演化计算，因为没有更好的替代品。专家系统？？？和其他对人类智力的模拟，不能准确预测以采取措施，不robust。挑战是不能因为我们选择的倾向限制了演化计算解决问题的潜力。
2 Possible applications of evolutionary computation
2.1 Introduction
Applications of evolutionary computation (EC) fall into a wide continuum of areas. For convenience, in this chapter they have been split into five broad categories:
• planning
• design
• simulation and identification
• control
• classification.
These categories are by no means meant to be absolute or definitive. They all overlap to some extent, and many applications could rightly appear in more than one of the categories.
A number of bibliographies where more extensive information on EC applications can be found are listed after the references at the end of this chapter.
2.2 Applications in planning
2.2.1 Routing
Perhaps one of the best known combinatorial optimization problems is the traveling salesman problem or TSP (Goldberg and Lingle 1985, Grefenstette 1987, Fogel 1988, Oliver et al 1987, Mu ̈hlenbein 1989, Whitley et al 1989, Fogel 1993a, Homaifar et al 1993). A salesman must visit a number of cities, and then return home. In which order should the cities be visited to minimize the distance traveled? Optimizing the tradeoff between speed and accuracy of solution has been one aim (Verhoeven et al 1992).
最知名的问题是旅行商问题，旅行商必须访问所有城市然后回家，求最小化总距离的访问顺序。在速度和访问路径不重复中权衡。
A generalization of the TSP occurs when there is more than one salesman (Fogel 1990). The vehicle routing problem is similar. There is a fleet of vehicles, all based at the same depot. A set of customers must each receive one delivery. Which route should each vehicle take for minimum cost? There are constraints, for example, on vehicle capacity and delivery times (Blanton and Wainwright 1993, Thangia et al 1993).
一个旅行商问题的扩展是多个商人，在一个车队中。一组顾客每个人收到一个货物，每个车（商人）应该怎么走花费最少。
Closely related to this is the transportation problem, in which a single commodity must be distributed to a number of customers from a number of depots. Each customer may receive deliveries from one or more depots. What is the minimum-cost solution? (Michalewicz 1992, 1993).
和这个问题相似的是交通运输问题，一种货物必须从很多个仓库送到很多顾客手上。每个顾客可能从一个或多个仓库收到货，求最小花费的解。
Planning the path which a robot should take is another route planning problem. The path must be feasible and safe (i.e. it must be achievable within the operational constraints of the robot) and there must be no collisions. Examples include determining the joint motions required to move the gripper of a robot arm between locations (Parker et al 1989, Davidor 1991, McDonnell et al 1992), and autonomous vehicle routing (Jakob et al 1992, Page et al 1992). In unknown areas or nonstatic environments, on-line planning/navigating is required, in which the robot revises its plans as it travels.
测试机器人路线是另一个路线规划问题。路径必须可行和安全并且没有碰撞。例如决定机器人到达目的地关节的活动，在未知或者变化的环境中，需要实时规划。
