At Robomech2018 in Kitakyushu, my two students presented following contents.
“Continuous Learning using Fractal Reservoir Computing”
“Learning of Correlation in Autonomous Decentralized Multi-Agent with Individual Objectives”
The first presentation consists of the following.
- Purpose: to mitigate the learning problem using neural network, catastrophic forgetting
- Idea: to utilize the modularity of fractal network
- Method: to facilitate the activation of each module by adding an appropriate task input corresponding to each module
The second presentation consists of the following.
- Purpose: to achieve completely decentralized autonomous systems for multi-agent reinforcement learning (MARL)
- Idea: to propose a new problem setting, bottom-up MARL, where individual agent has own reward
- Method: to determine the degree of cooperation according to the learned correlation between rewards of other agents and own states