Distributed autonomous multi-agent reasoning and classification systems have been thought of to be the basis of intelligence and have wide applications in the space of operational intelligence in closing the loop between sensing, analytics, and actions. This paper targets multi-agent systems that employ rulebased logics (i.e., rules that determine the output/response of an agent depending on the range of the input values) with pre-defined rules to accurately perceive the environment, and provide associated reactions. Such rule-based systems do not perform well in scenarios, where human generated rules cannot adapt to dynamic variations in the data distribution arising due to dynamic changes in the environment, especially if data dimensionality is very high. Examples of such scenarios exist wherever the sensed data arrives from the physical world - such as weather data, physical sensor data, human behaviour controlled data, etc. Clearly, to meet the adaptivity requirements of such scenarios we require the agents to possess adaptive reasoning capability such that they can adapt the underlying rules with respect to the changing environment. Developing such adaptive agents requires the developer to additionally possess considerable expertise of state-of-the-art machine learning techniques, apart from possessing knowledge of the agent's target domain. To address the above issues, we automate the process of development and deployment of adaptive agents. We present the fundamental design concepts behind the development of SynAdapt: a new adaptive meta-learning based multi-agent synthesis framework, that automates the synthesis of adaptive multi-agent systems from high-level user specifications. SynAdapt provides the following key features: a) Automated synthesis and deployment of adaptive agents from high-level user specification, b) Agents synthesised by SynAdapt can select a learning strategy that is particularly suited for given user specifications and input dataset, and c) Agents synthesised by SynAdapt can leverage adaptive ensemble learning techniques to deal with concept drift.
[1]
Subhajit Sidhanta,et al.
An ad-hoc distributed execution environment for multi-agent systems
,
2013,
2013 Fifth International Conference on Communication Systems and Networks (COMSNETS).
[2]
Geoff Holmes,et al.
MOA: Massive Online Analysis
,
2010,
J. Mach. Learn. Res..
[3]
Geoff Holmes,et al.
New ensemble methods for evolving data streams
,
2009,
KDD.
[4]
Subhajit Sidhanta,et al.
Managing a Cloud for Multi-agent Systems on Ad-Hoc Networks
,
2012,
2012 IEEE Fifth International Conference on Cloud Computing.
[5]
Michel C. Desmarais,et al.
Improving matrix factorization techniques of student test data with partial order constraints
,
2012,
UMAP.