Cultural Algorithms. Robert G. Reynolds
Чтение книги онлайн.

Читать онлайн книгу Cultural Algorithms - Robert G. Reynolds страница 13

Название: Cultural Algorithms

Автор: Robert G. Reynolds

Издательство: John Wiley & Sons Limited

Жанр: Программы

Серия:

isbn: 9781119403104

isbn:

СКАЧАТЬ On pressing the blue “Start Run” button, the system will begin processing with the given attributes. A right clicking within the visual ConesWorld display window will allow a drag interface to move the visuals, and a mouse wheel can be used to zoom in and out. The left and right mouse buttons pressed in unison can be used instead of the mouse wheel to zoom. Different attributes of the run can be viewed via the tabs located at the bottom of the screen.

      Simphony is part of a suite of software known as the Repast Suite. This software is distributed by the Argonne National Laboratory. Repast Simphony is an agent‐based modeling software designed to organize, monitor, and record the movements of agents in a given environment.

      The system works as a shell, in that Simphony first executes its own visualization system, and then executes the provided environmental code within its own framework. Simphony also provides its own feedback for errors in code running within its environment for the purpose of debugging. It should be noted that errors in environmental code will not propagate out to the Java environment being used to run Simphony, so users should be aware of locating the feedback within Simphony itself.

      The Cultural Algorithm's implementation within the CAT system relies on a collection of Knowledge Sources, which can influence the behaviors of the agents as they work toward optimization. Each Knowledge Source analyzes the data in a different way, trending toward either widespread explorative style or tightly focused exploitative style. As the knowledge sources learn more about the problem being analyzed, their suggestions can improve performance resulting in more agents being influenced by that particular Knowledge Source. There are five Knowledge Sources used by the CAT system: Topographical, Normative, Domain, Historical, and Situational.

      The Topographical Knowledge Source influences agents to seek new possible feasible solutions by making predictions given the performance landscape of the scoring mechanism. By analyzing trends in certain regions, it can then make suggestions where it is likely that these trends may reach a maximum and influence agents to move toward those locations. Because it sends agents into unexplored areas based on its predictions, it falls into the Explorative category.

      In the CAT System, Topographical Knowledge takes the shape of a tree made up of the cells that contain information about the problem's topography. The topographical map can be composed of N‐dimensions, based on the number of variables within the given problem, plus the fitness function. For example, as seen in Figure 2.3, ConesWorld is a three‐dimensional topographical map, with the X and Z variables for location, and the Y value representing the fitness at a given location denoted by X and Z. The domain of X and Z is limited between the values of 0 and 250 for each of them, resulting in an area of 62 500 units for the given problem domain.

      The Normative Knowledge source experiments with the extension of norms. As the system progresses, clear standards begin to emerge that lead toward likely optimal solutions. If a new, greater optimal solution is found that exceeds the standards, then the knowledge source will readjust its constraints to encompass the new optimal solution.

      In the CAT System, the Normative Knowledge is represented by restricted ranges within the available ranges of the inputs. These ranges are adjusted over successive steps, with less optimal subsections of the ranges on the edges of the constraints discarded, and the intersections of these ranges of value are where all of the predictions from the Normative Knowledge source will be found. In ConesWorld, this can be visualized as two rectangular subsections of the topography that intersect, each subsection representing a single variable of the domain, within which agents will be placed.

      The Domain Knowledge is reflective of the world in which the problem itself exists. For example, a stock market program would be aware of seasonal trends, a system that works to determine optimal speed around a track would need to understand feasible speed around a curve, and a system that analyzes ConesWorld would need to understand the nature of the sloping topography. It examines the relationship between the objects that exist in a problem domain and how their interactions can lead to solutions.

      In the CAT System, the Domain Knowledge of which the system is aware is the nature of a cone. Through the use of a collection of agents and the slope of the area the agent is placed at, they can attempt to calculate where these different slopes would place a possible peak to be sought and a valley to be avoided. The more widely spread the agents are, the more inaccurate the prediction, yet the more tightly packed the agents, the more likely they are to be trapped in a suboptimal local maximum.

      To move the agents, the knowledge system influences them to gradually advance via a set step size in the direction that their current location's slope indicates, or along the direction indicated by the overall group's slope with regards to how many agents' slope directions point toward a common point. This is repeated for each dimension in the problem domain, with a small element of randomization included with the step size.

      When the agents are spread out, the Domain Knowledge acts in an explorative manner through the combined efforts of all of its agents. When they converge on a given point, it acts in an exploitative manner and agents can even become trapped if they all converge on a suboptimal point with a slope of zero at the peak of a cone.

      For this reason, the Domain Knowledge acts as a buffer between the Exploitative and Explorative Knowledge sources. It adapts to the nature of the problem domain via exploration, focuses in via exploitation, can then become trapped due to suboptimals, and then return to its explorative stage when another knowledge source provides it with fresh knowledge. It explores the possibilities of the system's governing rules and exploits them as well.

      Historical (or Temporal) Knowledge exploits past knowledge to guide agents in future predictions. By analyzing successful agents of the past and making cautious refinements on their variables, it is possible to predict a new optimal solution near to a high‐ranking solution of the past.

      To achieve this in the CAT system, a population of successful agents is recorded, in addition to a drift variable. The drift is the tolerance within which an agent must variate from the past recorded agent to be deemed worth the attempt. Even though this knowledge source exploits past successes, its drift ensures that it will not simply repeat the exact scenario of a past success for any given agent.

      As the CAT system also has a dynamic update for the landscape, the Historic Knowledge source has the ability to update itself in this situation as well. In a typical run, the historic record is kept and pruned as some past achievements are outpaced by a significant degree. The drift of the Historic Knowledge source can allow the prediction of a newly created optimal on the dynamic landscape if the new optimal location is a slight variation of a previous historic scenario.

      The most exploitative of all of the knowledge sources, СКАЧАТЬ