Home > Research > Publications & Outputs > Autonomous Machine Learning (ALMA)
View graph of relations

Autonomous Machine Learning (ALMA): generating rules from data streams

Research output: Contribution in Book/Report/ProceedingsChapter (peer-reviewed)


Publication date19/09/2011
Host publicationProceedings of the Special International Conference on Complex Systems, COSY-2011: 16-19 September 2011
Place of PublicationOhrid, FYR of Macedonia
Number of pages8
<mark>Original language</mark>English


In this paper the well known problem of automatically generating tractable models (for example, but not limited to fuzzy rules) from data by learning will be given a new impetus by introduction of the concept of Autonomous Learning MAchines (ALMA). This concept increases the level of automation and autonomy of the process of model identification and design by reducing the need for human intervention in parts of the problem where this is traditionally done manually and off-line such as model structure identification. The proposed concept is generic and is not limited to fuzzy rule based (FRB) systems or neuro-fuzzy (NF) systems, but is also applicable to hidden Markov models (HMM), decision trees between others. For the specific case of FRB and NF models which are considered in this
paper, the membership functions (MF) are automatically generated from (and represent) the true data distribution using kernels and data clouds and recursively estimated relative data density. In this paper we propose an original method for evolving clouds from data based on the well known mean-shift approach and we propose an evolving version of it which we call evolving local means (ELM). The proposed ALMA can be used as a basis of software algorithms and agents and hardware devices in a wide range of problems in industry, defense, security, space exploration, robotics, human behavior analysis, assisted living, etc.