| Research Project | |
| MiNet Seekey M cubed Flame iDSS |
|
| Flame Towards Flexible Agent Mechanism for the Dynamic Environment |
||
 |
||
| member | Hiromitsu HATTORI，Okina YAMAGUTI，Takahumi YAMAYA，Yujiro HUKAGAYA | |
|
||
| keyword | Multiagent Systems, Negotiation/Argumentation, Constraint Satisfaction Problems, Soar, Auction, Probabilistic Reasoning, Reinforcement Learning | |
|
||
| mission | Research project "Flame" elaborates to enhance Multi-agent Systems by constructing next generation of agent technologies. Especially, we try to realize an ability to flexibly handle the dynamics of environment. | |
|
||
| outline |
The member of Flame try to investigate and develop the
sophisticated agent negotiation/coalition formation methods
for deeepening of Multi-agent Systems. In the dynamic
environment, which includes vast amounts of uncertain
information (e.g., Internet), agents must flexibly reason,
negotiate, and make a decision based on the uncertain
information, and they must cosider with the dynamics
of the environment. Therefore, we have to develop/propose
the effective methods for constructing next generation of
agent technologies. Currently, we focus the following subjects.
Scheduling has been an important research field in Artificial Intelligence. Because typical scheduling problems could be modeled as Constraint Satisfaction Problems‾(CSP), several constraint satisfaction techniques have been proposed. In order to handle the different levels of importance of the constraints, solving the problems via Valued Constraint Satisfaction Problems‾(VCSP) is an promising approach. However, there exists the case where an unexpected events which might require a sudden change in the obtained schedule, i.e., the case with dynamic changes in scheduling problems. In this research, we describe such dynamic scheduling problem as Dynamic Valued Constraint Satisfaction Problems‾(DyVCSP) in which constraints would changes dynamically. Generally, it is undesirable to determine vastly modified schedule even if re-scheduling is needed. A new schedule should be close to the current one as much as possible. In order to obtain stable solutions, we propose the methodology to maintain pieces of the current schedule via the use of temporal soft constraints, which is explicitly penalizing the changes from the current schedule. In this research, we focus on the nurse scheduling for applying our method. The dynamic soar-rule generator based on reinforcement learning technique We present a Soar architecture agent using reinforcement learning that has two features: 1) making a plan based on the Soar architecture, and 2) resolving an impasse with reinforcement learning. A Soar agent normally makes a plan with production rules in recognition memory. When a Soar agent cannot solve a problem with its recognition memory, it generates an impasse. To resolve an impasse, a Soar agent autonomously starts a reinforcement learning unit and generates new production rules from the result of the unit. The agent generates production rules with two steps: 1) extracting association rules from a Q-function that is learned with a reinforcement learning method, and 2) deriving production rules from association rules with EBL(Explanation-Based Learning) like method. The second step uses a user ’s background knowledge to reduce the number of production rules, and to facilitate comprehending what rules has been derived. Generated production rules are stored in the recognition memory. We have implemented a problem solver for the container task environment to show how effectively our Soar agent can work. |
|
|
||
| publications |
|
|
 Shintani Lab. 2004 |
||