Flentge, Felix ; Schneider, Christoph ; Jung, Tobias ; Metz, Sascha ; Deusser, Robert (2005)
Team Description Mainz Rolling Brains 2005.
Anderes, Bibliographie
Kurzbeschreibung (Abstract)
This year we will introduce a whole new decision layer design based on gen- eral principles for coordinated multiagent decision making. In last year's Mainz Rolling Brains 3D agent [1] we used the same agent design as in our 2D agent [2{4]. On the decision layer we had several modules for different tasks as passing or dribbling. These modules rated the usefulness of their particular actions de- pending on the current situation. Then the module with the highest rating was called to execute its action. We identified two main problems with this approach: First, the coordination between these modules turned out to be quite difficult. To obtain a different agent behavior, changes in several modules were necessary. Also, it was not very clear which module is used in which situation. Second, the switch from 2D to 3D simulation means a considerable increase in the complexity of the agent control. In particular, the handling of the ball got a lot more difficult than before and so now an additional couple of steps is necessary to accomplish the basic tasks as passing or dribbling. This has far reaching consequences. For example, in 2D it was sufficient to think about the pass partner when the agent was in the position to kick the ball because it could kick in every direction quickly. In 3D the agent has to choose its pass partner many simulation cycles before it reaches the ball because it can only kick in the direction from the agent towards the ball. So, in order to play a pass towards a teammate, it has to approach the ball in a certain way. While approaching the ball, the pass partner may also change its position which would make further adaptations necessary. Our answer to these problems is to introduce new methods for making de- cisions and executing actions. The basic idea is to use team actions which are valid for the whole team and facilitate the coordination of the agents. Based on these team actions we will have a clear structure for choosing the appropriate individual actions using explicit goal functions which rate the usefulness of the outcomes of the team actions. This general approach to coordinated multiagent decision making is described in the next section. In section 3 we will give a short overview of the overall agent design and give some hints how the general approach is realized in the decision layer. Section 4 deals with the learning of a simple skill: a simulated two-wheeled robot learns to approach a moving ball in order to kick it towards a given target. We conclude with section 5 where we sum up the current status and describe our plans for the future.
| Typ des Eintrags: | Anderes |
|---|---|
| Erschienen: | 2005 |
| Autor(en): | Flentge, Felix ; Schneider, Christoph ; Jung, Tobias ; Metz, Sascha ; Deusser, Robert |
| Art des Eintrags: | Bibliographie |
| Titel: | Team Description Mainz Rolling Brains 2005 |
| Sprache: | Deutsch |
| Publikationsjahr: | 2005 |
| Kurzbeschreibung (Abstract): | This year we will introduce a whole new decision layer design based on gen- eral principles for coordinated multiagent decision making. In last year's Mainz Rolling Brains 3D agent [1] we used the same agent design as in our 2D agent [2{4]. On the decision layer we had several modules for different tasks as passing or dribbling. These modules rated the usefulness of their particular actions de- pending on the current situation. Then the module with the highest rating was called to execute its action. We identified two main problems with this approach: First, the coordination between these modules turned out to be quite difficult. To obtain a different agent behavior, changes in several modules were necessary. Also, it was not very clear which module is used in which situation. Second, the switch from 2D to 3D simulation means a considerable increase in the complexity of the agent control. In particular, the handling of the ball got a lot more difficult than before and so now an additional couple of steps is necessary to accomplish the basic tasks as passing or dribbling. This has far reaching consequences. For example, in 2D it was sufficient to think about the pass partner when the agent was in the position to kick the ball because it could kick in every direction quickly. In 3D the agent has to choose its pass partner many simulation cycles before it reaches the ball because it can only kick in the direction from the agent towards the ball. So, in order to play a pass towards a teammate, it has to approach the ball in a certain way. While approaching the ball, the pass partner may also change its position which would make further adaptations necessary. Our answer to these problems is to introduce new methods for making de- cisions and executing actions. The basic idea is to use team actions which are valid for the whole team and facilitate the coordination of the agents. Based on these team actions we will have a clear structure for choosing the appropriate individual actions using explicit goal functions which rate the usefulness of the outcomes of the team actions. This general approach to coordinated multiagent decision making is described in the next section. In section 3 we will give a short overview of the overall agent design and give some hints how the general approach is realized in the decision layer. Section 4 deals with the learning of a simple skill: a simulated two-wheeled robot learns to approach a moving ball in order to kick it towards a given target. We conclude with section 5 where we sum up the current status and describe our plans for the future. |
| Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik > Telekooperation 20 Fachbereich Informatik |
| Hinterlegungsdatum: | 31 Dez 2016 12:59 |
| Letzte Änderung: | 15 Mai 2018 12:01 |
| PPN: | |
| Export: | |
| Suche nach Titel in: | TUfind oder in Google |
 |
Frage zum Eintrag |
Optionen (nur für Redakteure)
 |
Redaktionelle Details anzeigen |
Drucken
Impressum