From Krzysztof Krawiec

Site: ExplanatoryModeling

New Computational Paradigms for Explanatory Modeling of Complex Systems

I'm coordinating the research project New Computational Paradigms for Explanatory Modeling of Complex Systems (Nowe Paradygmaty Obliczeniowe dla Konstrukcji Wyjaśniających Modeli Systemów Złożonych), granted by National Science Center, DEC-2011/01/B/ST6/07318, 8.12.2011-7.12.2014.

Keywords: complex systems, modeling, evolutionary computation, genetic programming, machine learning, feature selection and construction.

Project investigators

• Krzysztof Krawiec, principal investigator
• Roman Słowiński
• Szymon Wilk
• Wojciech Jaśkowski
• Tomasz Pawlak
• Bartosz Wieloch
• Marcin Szubert
• Paweł Liskowski
• Karolina Stanisławska
• Mateusz Nawrocki
• Jerzy Błaszczyński
• Robert Susmaga
• Piotr Zielniewicz

Scientific Goal

This project is intended to combine in a novel way selected concepts from branches of artificial intelligence (AI) and computational intelligence, primarily from machine learning and evolutionary computation, to design new methods and algorithms capable of constructing explanatory models for data produced by complex systems.

The primary objective of this project is to elaborate a new paradigm for modelling and analysis of complex systems that combines the advantages of the data-driven search, characteristic for machine learning, with the performance-driven search, characteristic for evolutionary computation. This will boil down to designing, implementing, and testing a family of new algorithms capable of producing models of complex systems that have desirable properties. In particular, this project is devoted to:

Models that:

1. Are expressed in legible ways (preferred symbolic representation) and plausibly explain the system,
2. Generalize well to testing data (if available).

Methods that:

1. Are prepared to handle large volumes of training data,
2. Produce high-quality (plausible) models in reasonable time,
3. When needed, autonomously introduce additional variables that describe internal states of the complex system, in order to make the model more accurate and/or comprehensible,
4. Discover selected properties of investigated complex systems, in particular modularity and hierarchy, and benefit from that fact (in terms of the quality of models and runtime),
5. Are able to model multiple dependencies between the components of the system (e.g., variables), particularly in the case when many dependent variables are present.

Importance

In this proposal we admit the definition of complex system that is widely accepted in the computational intelligence community. We define it as a system that is composed of many (often a large number of) components and exhibits some emergent behavior(s), i.e., behavior(s) that cannot be explained by considering the constituting components separately (Mitchell 2009). Other features often displayed by complex systems include:

• nonlinearity and non-monotonicity of dependencies between components (e.g., variables),
• large volumes of produced training data (even in a single episode that results in a single example),
• large numbers of possible system behaviors (of which the training data are only a limited sample),
• indeterminism.

Indirectly, these features determine the class of problems we would like to approach and the class of training data collections we would like the developed methods to be able to handle.

The central tenet of this project is that, given the recent progress in AI and computational intelligence, and the increasing amount of available computing power, it is both scientifically legitimate and meaningful to pursue the search for new methods that can efficiently discover models that explain complex systems while making moderate use of abstraction and domain-specific knowledge.

Research plan

The proposed project will be carried out within a few main tasks detailed below.

• Task 1. Modelling complex systems using genetic programming, rule systems, and their hybrids.
• Task 2. Feature selection and construction methods oriented towards modelling of complex systems.
• Task 3. Extensions involving exploitation of selected properties of complex systems.
• Task 4. Modelling of real-world complex systems.