Project/Area Number |
23KF0108
|
Research Category |
Grant-in-Aid for JSPS Fellows
|
Allocation Type | Multi-year Fund |
Section | 外国 |
Review Section |
Basic Section 61040:Soft computing-related
|
Research Institution | The University of Tokyo |
Principal Investigator |
池上 高志 東京大学, 大学院総合文化研究科, 教授 (10211715)
|
Co-Investigator(Kenkyū-buntansha) |
CROSSCOMBE MICHAEL 東京大学, 大学院総合文化研究科, 外国人特別研究員
|
Project Period (FY) |
2023-07-26 – 2026-03-31
|
Project Status |
Granted (Fiscal Year 2023)
|
Budget Amount *help |
¥2,100,000 (Direct Cost: ¥2,100,000)
Fiscal Year 2025: ¥600,000 (Direct Cost: ¥600,000)
Fiscal Year 2024: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2023: ¥800,000 (Direct Cost: ¥800,000)
|
Keywords | Collective intelligence / Collective dynamics / emergent behavior / neuroevolution / simulation |
Outline of Research at the Start |
The main goals of this project are: 1) To study the impact of spatial constraints and information bottlenecks on collective behaviour, to improve our understanding of the necessary conditions for collective intelligence; 2) To develop a new framework for the evolution of collective behaviours which incorporates constraints observed in living systems; 3) To demonstrate the validity of this approach by evolving new collective behaviours in simulation.
|
Outline of Annual Research Achievements |
This year we have developed a simulation environment to facilitate the evolution of artificial neural networks (neuroevolution) to infer a general model of collective behaviour exhibited by a target living system’s observed dynamics. In this case, our simulation environment uses collective behaviour data obtained from recordings of real ant colony dynamics. This is then used to automatically evolve neural network topologies which attempt to reproduce the general behaviours observed in living systems. This framework can be applied to any living system whereby the observable dynamics can be represented as a time series of 2-dimensional positional data.
So far, we have submitted an extended abstract about the simulation environment which we intend to make available to the research community.
|
Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
At the ALIFE2023 conference (hosted by Prof. Iizuka at Hokkaido University), our paper on the results of the information bottleneck approach was accepted for an oral presentation. Our overall research plan is making good progress.
|
Strategy for Future Research Activity |
From here, we are already investigating methods to evolve more complex neural network topologies which are suitable for capturing the complex dynamics of living systems. First, we will redesign the current reward function to ensure that the general dynamics of the living system being modelled are accurately reproduced for learning. Then, extending existing algorithms for topological neuroevolution (NEAT, WANNs), we will evolve networks with more sophisticated internal architectures that might eventually evolve internal state representations. To do this, we will no longer stick to a strictly feed-forward architecture but instead allow the evolutionary process to use recurrent connections and additional activation functions in tandem. We expect this will produce more accurate dynamics.
|