Strategies and prediction in multi-agent systems

Chapter 17 - Strategies and Prediction in Multi-Agent Systems

One of the problems faced by soccer robots is how to find out the position of the robot itself and other robots on the field. A simple way to find out the robot's position is to use the odometry method. However, odometry is weak in accumulating position errors that reduce the accuracy of the moving robot's absolute position estimation and orientation. This paper presents a robot position data estimation system that is to be implemented on the ERSOW wheeled soccer robot. The robot can determine its position based on a unique landmark: an L-shaped line on the soccer robot field. We use a deep neural network method to recognize landmark L-shaped. Vision systems and deep learning inferences run on the Robot Operating System platform. After obtaining the distance of the robot to the L-shaped landmark, the robot's orientation and position relative to the field can be accurately determined based on the omnidirectional camera's perception. The results of the position estimation system in this study can be used to reduce position errors resulting from the odometry method. Based on the landmark L-shape recognition test results conducted on 641 datasets, the validation accuracy value is 0.806. The results of testing the robot position generated by vision obtained the largest error x about 2.32 cm and y about 1.99 cm.

Coordination of the multi agent system and ITS application to robot soccer

Coordination of multiple autonomous robots is a typical research problem in Multi-agent systems. The distributed agent models that represent the supervisory controllers of autonomous soccer robots are capable of communication, interaction and collaboration. By introducing the multi-agent based control architecture, the soccer robots perform seamless team cooperation to achieve the desired mission during competition. Due to the multi-soccer robot system is a discrete event dynamic system with concurrent and asynchronous characteristics, the distributed agent-oriented Petri net (DAOPN) is selected in this paper to model and control the behaviors of the multi-soccer robot systems. The agent models proposed in this work consist of the coordination model and the skill oriented robot behavior models. Especially, the agent based communication places enable the distributed agent models to be communicated directly. In this manner, the DAOPN based agent models can act as the robot controllers and coordinator to perform run-time controls of the soccer robot team. Therefore, the proposed approach not only resolves problems of the multi-robot coordination, but also provides a feasible way to overcome the complicated mathematical functions or rule bases. Finally, this system is tested using four soccer robots, and the results verify the proposed DAOPN based controller. Index Petri nets, agent-based simulation and modeling, cooperation and coordination among agents.

Distributed play-based approaches have been proposed as an effective means of switching strategies during the course of timed, zero-sum games, such as robot soccer. However, play-based approaches have not yet been rigorously evaluated in a full robot soccer scenario. In this paper, we perform an extensive empirical analysis of play effectiveness with teams of robots. We show that different plays have a significant effect on opponent performance in real robot soccer games. In our analysis we further realized the problem of distributed play recognition: classifying the strategy being played by the opponent team. Play recognition in real robot soccer is a particularly challenging problem because our observations are only “incidental” - that is, the primary task of our team is to play soccer, not to explicitly observe members of the other team. Despite these challenges, we achieve high classification accuracy in the robot soccer domain. To achieve high accuracy, our team maintains a history of joint observations including team positions, opponent positions, and ball positions and utilizes hidden Markov models to recognize opponent plays.KeywordsHide Markov ModelBall PositionObservation SequenceRobot TeamRobot SoccerThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Multiagent systems in which independent software agents interact with each other to achieve common goals, complete distributed tasks concurrently under autonomous control. Agent communication has been shown to be an important factor in coordinating efficient group behavior in agents. Most research on training or evolving group behavior in multiagent systems used predefined agent communication protocols. Designing agent communication becomes a complex problem in dynamic and large-scale systems. In order to solve this problem, in our previous research we proposed a method applying genetic programming techniques, in particular Automatically Defined Function Genetic Programming (ADF-GP) (K. Mackin and E. Tazaki, 1999), to allow agents to autonomously learn effective agent communication messaging. For this research we take this approach further and combine training of the agent behavior as well as the communication protocol. By training both behavior and communication we expect to further optimize the system performance. A software simulation of a multiagent transaction system is used to observe the effectiveness of the proposed method.

Multiagent systems, in which independent software agents interact with each other to achieve common goals, complete concurrent distributed tasks under autonomous control. Agent Communication has been shown to be an important factor in coordinating efficient group behavior in agents. Most researches on training or evolving group behavior in multiagent systems used predefined agent communication protocols. Designing agent communication becomes a complex problem in dynamic and large‐scale systems. In order to solve this problem, in this paper we propose a new application of existing training methods. By applying Genetic Programming techniques, namely Automatically Defined Function Genetic Programming (ADF‐GP), in combination with pheromone communication features, we allowed the agent system to autonomously learn effective agent communication messaging for coordinated group behavior. A software simulation of a multiagent transaction system aiming at e‐commerce usage will be used to observe the effectiveness of the proposed method in the targeted environment. Using the proposed method, automatic training of a compact and efficient agent communication protocol for the multiagent system was observed.

When planning the soccer robot path at present, a two-dimensional map is used mainly to optimize the path of the soccer robot's operating field. However, because the two-dimensional robot is selected, which can only utilize the data of plane about a specific environment, the data information of the mobile soccer robot cannot be gathered, impacting greatly the completion of the plan of the robot path. The path planning of the soccer robot is conducted using the quantum genetic algorithm so that the problem can be dealt with. On the premise that there is a lack of full consideration of the accurate motion path between the two points, the inertia weight is dynamically adjusted to overcome the disadvantage of premature convergence of the traditional quantum group algorithm, which can make the weight of the quantum genetic algorithm controllable with adaptability so that the problem of the extremely slow convergence speed of a single quantum in the quantum group and the high group dispersion can be solved. Furthermore, the selected quantum genetic algorithm can realize the real-time update of the soccer robot's position, effectively increase the trajectory change of the particle swarm movement, and to a certain extent effectively increase the search ability and convergence effect of the particle swarm in the global scope, so as to ensure that this algorithm can be more effective than the traditional path planning algorithm in terms of global search ability and convergence speed in the path planning of the soccer robot and has a higher use value.

A novel scale-invariant feature transform (SIFT) algorithm is proposed for soccer target recognition application in a robot soccer game. First, the method of generating scale space is given, extreme points are detected. This gives the precise positioning of the extraction step and the SIFT feature points. Based on the gradient and direction of the feature point neighboring pixels, a description of the key points of the vector is generated. Finally, the matching method based on feature vectors is extracted from SIFT feature points and implemented on the image of the football in a soccer game. By employing the proposed SIFT algorithm for football and stadium key feature points extraction and matching, significant increase can be achieved in the robot soccer ability to identify and locate the football.

The use of norms is widely accepted as an effective approach to control and regulate the behaviour of agents in multi-agent systems. Existing work on normative multi-agent systems has mainly focussed on how norms can influence the behaviour of agents by assuming that the agents' behaviours are perfectly monitored. In this paper we focus on monitoring mechanisms, propose different types of monitors, provide a logical analysis of monitors, study the relations between monitors and norms to be monitored, and finally explore computational aspects of norm monitoring.

This paper presents a benchmark for multiagent systems specific to the simulator Soccerserver 2D, an environment to develop teams of robotic soccer, providing metrics and evaluation procedures for multiagent organization schemes, more specifically, coalitions formation. This benchmark has considered a MAS with two main levels, at least: (i) individual level, where agents are implemented from requisites of a social structure and considering its individual capabilities (roles, skills, etc); (ii) a social level, where all the social aspects of the MAS are specified (organization, plans, goals, etc.) and where the individual level of each agent instantiates these social knowledge to act in the system. The method proposed here has applied at social level, once it measures the quantity and quality of coalitions that arise in the environment.

Since the quality of data affects the success rate ofdata mining and learning algorithms, it is always attempted to identify and remove the irrelevant and redundant information in a dataset. Robotic soccer is a multi-agent system in which agents play in real-time, dynamic, complex and noisy environment. Many parameters affect the result of shooting toward the goal and agents must response to variations in soccer field rapidly. Therefore it is impossible to use all features in scoring behavior. This paper selects dataset for effective features of scoring behavior simulated soccer agents, then compares the size of the trees and accuracy produced by each feature selection scheme against the size of the trees and accuracy produced by C4.5 with no feature selection method. Experimental results have shown that dimensionality reductions lead to operate faster and more effective learning algorithm in real-time simulated soccer agent.

The goal of the robotic soccer competition is to develop completely autonomous robotic agents that can cooperate to perform the desired tasks in an extremely dynamic environment. The approach is to train the robots, using various combinations of learning algorithms, i.e., the robots are to initially have no understanding of their environment. The robots are rewarded for positive contributions to the team, and in turn learn better performance through multiple sessions. The majority of learning will take place in simulation and then transferred to real physical mobile robots. The concept of tropism-based control architecture is introduced that not only allows for the evolution of cooperative strategies, but also obtains the acquired knowledge in a format that is easily comprehensible by humans. The results of many generations of simulated evolution are presented, accompanied by the game results and fitness characteristics.

The main purpose of this work is the recognition of soccer team formations by considering a dynamic structural analysis. Traditionally, the recognition of team formations is carried-out without taking into account an expressive representation of relations between players. This kind of approaches are not able to manage the constant changes occurring in the soccer domain, which results in an inefficient way of recognizing formations immerse in dynamic environments. It is presented in this work an efficient model to recognize formations based on a representation that takes into account multiple relations among defender, midfielder and forward players. The proposed model has been tested with different teams in off-line mode showing that it is able to recognize the different main formations used by a team during a match.

For one agent, the proper selection of its own actions or strategies is one key to the multi-agent system. A new method is presented to select actions cooperatively for soccer robots, which is regarded as a benchmark of multi-agent system. The perfect character of artificial immune system is made full use to improve the strategy of the soccer robot, dealing with information of the environment as antigen, certain action as antibody. Like biological immune system, the robot selects action (antibody) according to its own surroundings (antigen) to make robots complete their mission properly and cooperatively. Experiment results indicate the feasibility and effectiveness of the method.