Home(English)

Home(Japanese)

My papers Abstraccts

Do not renewed since 1999. See List

To improve the prediction ability of an intelligent system, information should be decomposed into reusable parts, each of which contains a rule. We firstly propose a new criterion called Matchability suitable for the decomposition of real world information, and then propose a decomposition method. We developed a search algorithm for decomposition. Simulation demonstrates that the algorithm can decompose mixed situations. This technology is effective for pre-processing of data analysis and pattern recognition.

KW: Self-organization, Data analyses

[List/Abstract/Paper(English/Japanese)/Poster]

Patients who have part of their frontal lobe injured exhibit the phenomenon of perseveration. If we define perseveration as the result of losing the ability to take action appropriate for changing situations, i.e. the loss of attention learning, we can make a neural network model that directs attention to the modality appropriate for a particular situation. We have pointed out the importance of negative reinforcement learning, which takes place when the evaluation score decreases. The computer simulations have proven that the results obtained with the proposed model as representative of attention learning in the brain agree well with the results of psychological tests.

KW: frontal lobes, perseveration, Wisconsin Card Sorting Test(WCST), negative reinfoecement learning

[List/Abstract]

AU: Hiroshi YAMAKAWA, Hiroyuki OKADA, Nobuo WATANABE Tomoharu MOHRI
LN: Japanese
JN:16th Annual Conference of the Robotics Society of Japan (RSJ'98)
VN:Vol.2 pp.647-648, (2I21)
AB:

The mobile robot that has infrared sensor to all directions is simulated. The partial situation that detects near the wall is autonomously acquired by the situation decomposition technique based on matchability criterion.

KW: Situation Decomposition, Matchability, Mobile Robot, Learning

[List/Abstract/Paper]

We developed none-contract 3D measuring system using moire topographic method. Proposed system can measure 3D distance in the patch unit by detection moire interference fringes caused by a xenon light. By suing a xenon light, mobile robot's experiment can be done safely even if a person is by the side.

KW: Mobile robot, Range finder, Moire topographic method

[List/Abstract]

We propose the Cognition based InTelligent Transaction Architecture (CITTA) to realize autonomous learning function in real world intelligent systems. CITTA is pattern-based architecture to avoid symbol-grounding problem. CITTA is cognition-based architecture to acquire the information structure of facing environment. Computer simulation of tracking task demonstrates that system based on CITTA integrate knowledge which are distributed on agents.

KW:

[List/Abstract]

To realize real world autonomous learning function, we are developing distributed intelligent architecture called cognition based intelligent transaction architecture (CITTA). Main points for introduce learning ability is (1) Maintaining real world pattern information directly, (2) Strong formalization to reduce learning space. Firstly, we introduce a bidding mechanism on CITTA to combines distributed knowledge on each agent. The simulation of tracking task demonstrates the ability of system based on CITTA. Secondly, we developed attention mechanism on CITTA based on physiological knowledge. Proposed model can adapt to change of effective input modality using reinforcement learning and can explain the results of cognitive experiment called Wisconsin card sorting test (WCST).

KW:

[List/Abstract/Paper]

AU: Yamakawa, H., Takahashi, H, Suehiro, T.

LN: Japanese

JN: MACC'97..

VN:

AB

To realize real world learning we use multi agent system (called Agent network) which has reusability and genelalizability. Main point for introduce learning ability is (1)Maintaining real world pattern information directly. (2)Strong formalization to reduce learning space. We already applied agent network model for manipulator. In this paper we introduce contract mechanism to the agent network model, and show that distributed knowledge is integrated to accomplish simple chasing task.

KW:

[List/Abstract/Paper/Poster]

KW:Attention, Situation, Reinforcement Learning, Wisconcin card sorting test, Agent network, Matchabiliyt

[List/Abstract/Paper/Poster]

We propose the concept of "agent network" to realize flexible and unified integration of multi-agent intelligent robot systems. A hand-to-hand robot system developed on the network proved effectiveness of the method.

KW: .Agent network, Intelligent robot, System integration, hand-to-hand task

[List/Abstract]

The Active Intelligence Laboratory has been conducting the project named "Real World Adaptive Autonomous Systems." On the development of the real robot system, we adopt multi-module system approach for easy incremental scaling up of the sensor-motor integrated system. To make multi-module system more adaptive, the system should flexibly integrate distributed function and change its internal structure dynamically without semantics given by supervisor. For this purpose, our proposing agent network formalizes the element processes in each agent and communication form between agents, and mainly uses non semantic communication channel on the contrary traditional AI technique. We implemented an instance of the networks on workstations. A hand-to-hand robot system was developed on the agent network. It realized robust and flexible task execution.

KW:

[List/Abstract/Paper]

LN: English

JN: Proc.International Conference on Pattern Recognition (ICPR-96).

VN: vol.4, pp.123-127, Viena, Austria, 1996

AB

For effective recognition, a recognition structure that controls the information flow among the specialized processing modules should reflect the implicit correlation structure of the environmental input. Autonomous construction of a recognition structure will lead to extensive improve in the flexibility of the adaptive recognition system. For this purpose we propose a matchability-oriented feature selection that can collect highly correlated features at each local module. Conventional techniques, which collect features that are more independent, are not suitable. Matchability is a concept derived from the recognition functions of an adaptive intelligent agent (useful for action generation) and corresponds to the probability of input data items matching stored data items in the recognition system. Proposed algorithm changes the weights attached to each feature depending on the degree of matchability of each feature. This algorithm could select highly correlated features in simple simulation.

KW:

[List/Abstract/html]

TI: Multi-Sensor Fusion Model for Constructing Internal Representation using Autoencoder Neural Networks

AU: Yoshinori Yaginuma, Takashi Takashi and Hiroshi Yamakawa

LN: English

JN: Proc. International conference on neural networks(ICNN-96)

VN: (Scheduled)

AB

KW:

[List/Abstract]

TI:Informationintegration model with autoencoding neural network

AU:Kimoto, T., Yaginuma, Y. and Yamakawa, H.

LN:Japanese

JN:Trans. IEICE

VN:vol.J79-D-II, no.5, pp949-959, 1996.

AB:

 

KW:

Information integration, Sensor fusion, Neural networks, Hierarchie, Internal representation

[List/Abstract/]

TI:Feature Selection for Acquiring Internal Structure of Recognition System based on Matchability

AU:Yamakawa, H

LN:Japanese

JN: Technical report of IEICE

VN: vol. 96, no.41, PRMU96-12, pp.1-8, 1996

AB:

For effective recognition, a recognition structure that controls the information flow among the specialized processing modules should reflect the implicit correlation structure of the environmental input. Autonomous construction of a recognition structure will lead to extensive improve in the flexibility of the adaptive recognition system. For this purpose we propose a matchability-oriented feature selection that can collect highly correlated features at each local module. Matchability is a concept derived from the recognition functions of an adaptive intelligent agent (useful for action generation) and corresponds to the probability of input data items matching stored data items in the recognition system. We check this algorithm in simple artificial environment.

KW:

Learning internal representation, Pattern recognition, Neural networks, Multi-agent system, Unsupervised learning, Case based reasoning

[List/Abstract/Postscript/HTML]

TI:Arm Robot with Multi Antenna and Tactile Sensor

AU:Yamakawa, H. & Suehiro, N.

LN:Japanese

JN:13th RSJ ??????

VN: vol.3, pp.741-742. $BL@<#Bg3X!J@n:j!K (B, 1995.

AB:

For studying learning ability of agent assembled mechanism, we are developing sensor-based robot arm manipulator, because robot must protect itself during the exploring processes. We mount the four fingered hand on the robot arm, made shields, attached these shields to the robot arm, install tactile sensors and antennas as a proximity sensors on the shields.

KW:

Antenna, Arm robot, Hand robot, Multi-agent, Proximity sensor, Reinforcement learning, Tactile sensor

[List/Abstract/Postscript]

TI:UnsupervisedAcquisition of Hierarchical Internal Representation with AutoencoderNeural Networks

AU:Yamakawa, H., Kimoto, T., & Yaginuma, Y.

LN:English

JN:Int. Conf. on neural information processing (ICONIPS'95)

VN: vol.1 pp.213-216. Beijing, China, 1995.

AB:

For flexible adaptation to the large size real world information, an intelligent system should automatically acquire high-level internal representation. We propose a hierarchically connected autoencoder neural networks, which can learn high-level internal representation from raw sensor data. We demonstrate this with our model by taking two camera images of a target from different angles, and then restoring a three-dimensional portion of that target by learning with back-propagation algorithm for each network.

KW:

[List/Abstract/Postscript]

TI:Profiel of researchers

AU:Yamakawa, H.

LN:English

JN:RWCP NEWS

VN: vol.1, 1995.

AB:

 

KW:

[List/Abstract/html]

TI:Learning internal representation for multi sensor fusion using autoencoder neural networks

AU:Yaginuma, Y. Kimoto, T. & Yamakawa, H.

LN:Japanese

JN:11th Fuzzy system symposium

VN: FD2-4, pp.715-718. Ryukyu Univ., JAPAN, 1995.

AB:

For applying a system to real-world, it is important that the system can integrate from many kinds of information and obtain the essence with the system's purpose from that information. In this paper, at three-dimensional object recognition, we propose an autoencoder neural network model obtaining the essence for recognition with multi sensor fusion. We also verified the effectiveness of the proposed model by simulation exercises of a doll recognition problem, that uses camera image segmenting by that's hue and saturation.

KW:

Auto-encoder neural network, Sensor fusion, Internal representation, Object recognition

[List/Abstract/Postscript]

TI:Hierarchical Information Integration Model using Unsupervised Autoencoder Neural Networks

AU:Yamakawa, H. Kimoto, T. & Yaginuma, Y.

LN:Japanese

JN:11th fuzzy system symposium

VN: FD2-3, pp.711-714., Ryukyu univ. Okinawa, Japan, 1995.

AB:

For acquiring abstract internal representation from real world data, gradual learning and hierarchical architecture are essential. We propose hierarchical information integration model using auto-encode neural networks, which can learn high-level internal representation form raw sensor data. We demonstrate that the hierarchical model, which takes two camera images of one target from different view, learns to restore three dimensional potion of that target.

KW:

Auto-encoder neural network, Hierarchy, Internal representation, Stereo vision, Recognition

[List/Abstract/Postscript]

TI:A Neural Network-Like Critic for Reinforcement Learning

AU:Yamakawa, H., & Okabe, Y

LN:English

JN:Neural Networks

VN: vol.8, no.3, pp.363-373, 1995.

AB:

An adaptive agent that contains a reactive network and a critic that supervises that reactive network have been studied. Agent actions are generated in response to stimuli through the reactive network and they influence the ambient environment. The critic has a new learning algorithm which recursively enhances reinforcement signals from fixed reinforcement signals by interacting with the environment. The reactive network learns appropriate stimulus-action relations by reinforcement learning.

Computer simulation demonstrates that this neural critic is effective in environments where the concepts are embedded in a maze structure. We also suggest similarities between this critic model and the neural circuit in the human brain.

KW:

Reactive system, Neural network, Agent, Maze-like environment, Recursive structure, Amygdala

[List/Abstract]

TI:Pattern based intelligent system - Speculationon symbol grounding problem as view d from learning ability -

AU:Yamakawa, H.

LN:English

JN:Information Integration Workshowp (IIW-95)

VN: pp.94-103. Ibaraki, Japan.

AB:

Methods for realizing high-level intelligence based on symbols in an engineering way have been studied in the AI field for years, but those studies are not so successful in using these methods for handling pattern-intensive information from the real world. This is called the Symbol Grounding Problem, and is considered to be caused by the difference between pattern and symbolic information. In this paper, this problem is examined from the learning aspect, and it is shown that the problem resides essentially in the inability of autonomous high-level learning which involves generation of new internal representations in the portion of symbolizing pattern information. If so, it may be preferable to use only patterns for intelligent processing in order to construct systems which can deal with the real world. It is also pointed out here that with this method there remain such problems as the explosion of the amount of processing and difficulties in design.

KW:

Autnomous learning, Corelation, Observer, Learning of structre, Artificial intelligence, Real world

[List/Abstract/Postscript]

TI:Autoencoder Nwtwork Model Coping with Imcomplete Data

AU:Kimoto, T., Yaginuma, Y. & Yamakawa, Y.

LN:Japanese

JN: Technical report of IEICE

VN:vol.94, no.487, NC94-62, pp.17-24, 1995.

AB:

 

KW:

Autoencoder network, Internal representations, Incomplete data, 3D object recognitons

[List/Abstract]

TI:Computer comunity

AU:Yamakawa, H.

LN:Japanese

JN:RWCP Technical Report

VN:TR-95????, 1995

AB:

 

KW:

[List/Abstract/Postscript/PDF]

TI:Hierarchical Sensory-Motor Fusion Model with Neural Networks

AU:Nagata, S., Masumoto, D., Yamakawa, H., Kimoto, T.

LN:Japanese

JN:J. Robotics Society of Japan

VN:vol.12, no.5,pp.685-694,1994.

AB:

 

KW:

[List/Abstract]

TI:Active Data Selection and Subsequent Revision for Sequential Learning with Neural Networks

AU:Yamakawa, H., Masumoto, D., Kimoto, T., & Shigemi Nagata

LN:English

JN:World congress of neural networks(WCNN'94),

VN: vol.3, pp.661-666, San Diego, USA., 1994.

AB:

We propose a neural network system that sequentially obtains I/O sample data. The system selects useful sample data as training data, in what we call active data selection (ADS), and interpolates errors between training data and the network output, called subsequent revision (SR). ADS removes sample data if doing so only causes small errors. To speed up ADS, we ignore errors generated by the network and consider only those from SR.

We found that ADS steadily decreases errors and that SR gives suitable output, even if the neural network's learning is still not adequate. Simulation demonstrated the ability of the network to learn a sine function from sample data distributed unevenly in the input space.

KW:

[List/Abstract/Postscript]

TI:Sequential Learning with Neural Networks - Active Data Selection and Subsequet Revision -

AU:Yamakawa, H., Masumoto, D., Kimoto, T., & Shigemi Nagata

LN:Japanese

JN: Proc. Annual Conference of Japanese Neural Network Society(JNNS-93)

VN: pp.64-65. Kyusyu, Japan, 1993.

AB:

 

KW:

[List/Abstract]

TI:Hierarchical Sensory Information Processing Model with Neural Networks

AU:Kimoto, T., Masumoto, D., Yamakawa, H., & Shigemi Nagata

LN:English

JN:IEEE International conference on Robotics and Automation

VN:vol.1 pp.929-934. Atlanta, Jeorgia.

AB:

 

KW:

[List/Abstract]

TI: Active Data Selection and Subsequent Revision for Sequential Learning

AU: Hiroshi Yamakawa, Daiki Masumoto, Takashi Kimoto, Shigemi Nagata

LN:Japanese

JN: Technical report of IEICE

VN: vol.92, no.??, NC92-99, pp.33-40,1993.

AB:

We propose a neural network system that sequentially obtains I/O sample data. The system selects useful sample data as training data, in what we call active data selection (ADS), and interpolates errors between training data and the network output, called subsequent revision (SR). ADS removes sample data if doing so only causes small errors. To speed up ADS, we ignore errors generated by the network and consider only those from SR.

We found that ADS steadily decreases errors and that SR gives suitable output, even if the neural network's learning is still not adequate. Simulation demonstrated the ability of the network to learn a sine function from sample data distributed unevenly in the input space.

KW:

Neural networks, Sequential learning, Data selection, Training data, Interpolation, Back-propagation algorithm

[List/Astract/Postscript]

TI: Photoelectron spectra of liquid metals using synchrotron radiation

AU: Kakizaki, A., Niwano, M., Yamakawa, H., Soda, K., Ishi, T. and Suzuki, S.

LN: English

JN: J. Non-Crystalline Solids

VN: 117/118, pp.417-420. 1990.

AB:

The photeoelectron spectra of liquid and solid Bi were measured in UPS regin using synchrotoron radiation. It was found that the profiles of energy distribution curves are quite similar to those obatined in XPS regin and there is little difference in the spectral profile between liquid and solid phases. On melting, the density of states at the Fermi level increases more than twice as much as in the solid phase. In both liquid and solid phases, the photoelectron spectra change their profiles associated with the 5d core electron, which is attributes to the O4,5VV Auger electrons.

KW:

[List/Abstract]

TI: A UPS stydy of liquid and solid bismuth using synchrotron radiation

AU: Kakizaki, A., Niwano, M., Yamakawa, H., Soda, K., Suzuki, S. Sugawara, H., Kato, H., Miyahara, T., and Ishi, T.

LN: English

JN: J. Phys. F: Met. Phys.

VN: vol.18, pp.2617-2624, 1988.

AB:

The photoelectron spectra of Bi in both liquid and solid phases were measured at exitation energies between 18 and 60 eV using synchrotoron radiation. It was found that the profile of the energy distribution curves are quite similar to those obtaned in the XPS region and there is little defference inthe spectral profile between liquid and solid phases. On melting, the density of states at the fermi level increases more than twice as much as in the solid phase. In both liquid and solid phases, the photoelectron spectra change their profiles associated with the 5d core electron excitation, which is attributed to the O4,5VV Auger electrons.

KW:

[List/Abstract]