Typical fuzzy reinforcement learning algorithms take value-function based approaches such as fuzzy Q-learning in Markov Decision Processes and use constant or linear functions in the conclusion parts of the fuzzy rules. On the other hand, the policy-gradient approaches design policy functions directly and learn parameters included in the policy functions. Based on one of the policy-gradient approaches, a fuzzy reinforcement learning algorithm is proposed. This algorithm can deal with fuzzy sets even in the conclusion parts and also learn the rule weights of fuzzy rules. This paper's experiments show that the proposed learning algorithm is effective for a decision making problem of a soccer robot that plays in RoboCup Soccer Small Size League. After learning experiments with 30 scenes of a robot holding a ball, the robot control system learns a stochastic policy that agrees with that of human decision making in 25 of the 30 scenes.
In researches of multi-agent systems, evaluation method for multi-agent systems is one of important problems. Especially, evaluating agents, that behave on various environment, is difficult. In this paper, we analyze and clarify the relationship between map features and the evaluation of multi-agent systems in RoboCupRescue simulation. Through this paper, we can define evaluation indicators for each multi-agent system. And it makes it easy to compare one multi-agent system with the others on different environment. In other words, our final goal is to analyze evaluation values of some (rescue) agents'activity on an affected area and compare them. Besides, contribution of this paper is for complicated problem like disaster simulation, although traditional research is for simple problem like moving (or walking) simulation. In this point, we can regard this paper as a practical and important work. At first, we extract map features from a map. And then the features are quantified based on network analysis, urban engineering and so on. Road networks and building locations are adopted as such map features. Because rescue activity result depends on the road connectivity and building density. Therefore the features are defined as 5 classifications-building area, building element, location relationship among buildings, location relationship among buildings and roads, and density. And the score of the RoboCupRescue simulation is adopted as evaluation value of agents. Because therescue simulation includes many heterogeneous agents (firebrigades, ambulance team, and policeforce), integrated value is required to evaluate such agents. Before analyzing relationships between the map features and the evaluation values of the agents, it needs to remove multicollinearity from the defined the map features. And the results are analyzed based on partial correlation analysis. Through the analysis, we confirmed that the evaluation values of the multi-agent systems in RoboCupRescue simulation depend on the map features and that each algorithm of multi-agent systems has distinct dependency for the map features.
In this paper, we present a mechanical design of the holonomic omnidirectional vehicle using ball wheels for RoboCup Soccer Middle Size League (MSL). Most MSL team robots use the omni wheels, which have a lot of free rollers on the circumference of a wheel, to realize the omnidirectional motions. Omni wheels are easily implementable device, whereas they have drawbacks such as lack of moving stability due to the changes of friction or steps on the surface and their usable environment is limited. In order to play the games on a completely flat surface, we spread enough plates under the field carpet in MSL. Our aim is to overcome these drawbacks and develop actual prototypes for practical application not only for MSL. We evaluate the ability and accuracy of the prototype of omnidirectional vehicle using ball wheels.
In recent years, autonomous agents have been developed using statistical and probabilistic machine learnings together with deterministically optimized control. In this paper, a new decision making and motion generation method is proposed for adapting to uncertain environments. In the proposed method, we concentrate on passing in soccer, as a tactical group behavior, and understand how the optimization occurs in a group behavior from individual decision makings. In particular, we have quantified how people pass in plays by analyzing a video and tracking data of real soccer, and have constructed pass models with optimized parameters using logistic regression based on the analysis. As a result, our model predicted the next receiver with a high degree of accuracy by weighting positions of the players around the passer.
Robots were used to explore the World Trade Center (WTC) site after the September 11 terrorist strike, and the interior of the Fukushima Daiichi Nuclear Plant (FDNP) that was destroyed by the March 11 tsunami. Over the next several decades, robot design should be focused on use in unstable and dynamically changing environments such as those of WTC and FDNP. An important mission for rescue robots is making maps that can easily be understood by humans from information gathered on the changed terrain and victim locations in a disaster area. This automatic map-generating function is a popular topic for research on rescue robots. The performance of new and enhanced automatic map-generating functions must also be evaluated. In this paper, we propose a method for quantifying evaluation fields constructed for modeled uneven terrain and show the effect of a robot's movements on the performance of simultaneous localization and mapping methods.
Research on high level human-robot interaction systems that aims skill acquisition, concept learning, modification of dialogue strategy and so on requires large-scaled experience database based on social and embodied interaction experiments. However, if we use real robot systems, costs for development of robots and performing many experiments will be too huge. If we choose virtual robot simulator, limitation arises on embodied interaction between virtual robots and real users. We thus propose an enhanced robot simulator that enables multiuser to connect to central simulation world, and enables users to join the virtual world through immersive user interface. In this paper, we propose a framework of RoboCup@Home simulation based on such technology. We also show base technology and feasibility of the RoboCup@Home simulation system.
We have known the effectiveness for an algorithm of Ant Colony Optimization (ACO) and it has applied some applications. Real ants have communicated each other with pheromone materials and they search the shortest path from their colony to the place of foods. Researchers have reported the effectiveness of the algorithm for their applications. However there are some problems to resolve using this algorithm. One of them is a tradeoff problem between a convergence performance and a diversity of candidate for the optimized solution. On the other hand, from observations for real ants and their colonies, researchers in the field of biology have reported that there are two types of ants in the colonies. One of them is hardworking ants and another type is not hardworking ants. Then we have introduced different types of agents with ACO and we aim to generate stigmergy between agents. We have done some evaluation experiments in RoboCup Rescue Simulation System. From the results, we have considered the results between ACO with mono-agents and ACO with some types of agents. We have confirmed the effectiveness of our proposed method.
Visualization is a useful approach for knowledge discovery from databases and data mining, since it can support intuitive recognition of intrinsic structures of multi-dimensional observations. Fuzzy c-Varieties (FCV), which is one of the linear fuzzy clustering, performs local principal component analysis, and achieves the lower-dimensional local visualization of multi-dimensional data by using local principal components. Using squared Euclidean distances, however, the results of FCV are often sensitive to noise. In this paper, FCV is extended to a robustified version by using Alternative c-Means criterion, in which a modified distance measure is used based on a robust M-estimation concept. The proposed method is applied to a real world data set in order to construct local 2D-visualization, and the applicability of the visualization approach is investigated through knowledge discovery from the results.
In this paper, we suggest the ultrasonic-assisted technique of the thickness determination system on low anterior resection. We propose a technique for thickness determination of the intestine for low anterior resection using the ultrasound, whose center frequency is 15 MHz. Low anterior resection is one of the operative methods for rectum cancer. We determine the intestine thickness from ultrasonic waves using fuzzy inference. We performed the experiment using the biological phantom and the large intestine of the pig as a target object. As the results, we calculated the object thickness with the average absolute error of 5.67%.