ARTIFICAIL INTELLIGENCE DETAILED FRAMEWORK EXPLAINATION NEEDED

An Artificial Intelligence framework to promote the learning of complicated time series data models

Key words: artificial intelligence, logic programming, constraint programming, machine learning, bioinformatics

Background:

The systematic simulation and study of complex dynamical systems is the research area of this proposal (specifically in biological systems). The area of expertise of the MeForBio team (acronym for 'Formal Methods for Bioinformatics') of LS2N, one of the leading public research laboratories in digital sciences in France, is such a subject. The MeForBio team consists of 2 professors, 1 post-doctoral student and 2 computer science PhD students. It is heavily involved in local, national and international cooperation (particularly with the Tôkyô National Institute of Computer Science). Even if the application domain of this study is predominantly biology-oriented, there is no requirement for pre-requisites in life sciences. The key purpose of the current proposal is to develop creative machine learning algorithms that can capture the behaviour of large-scale complex systems and provide clarification.

The scientific impetus of this research project lies in the fact that it is now very easy to obtain a significant volume of time series data, with both I the spread of numerical instruments in every aspect of everyday life and (ii) the advancement in biology of New Generation Sequencing (NGS) methods. A key problem here is to apply a sense to these results, i.e. to construct relevant models (a challenge that can no longer be built by hands alone) that are both meaningful and predictive enough (for the researcher to have a deeper understanding of the processes at stake). To maximise one's comprehension of a targeted framework, it then becomes important to be able to link the time series data with models. This means that from the input data, we need to be able to learn the model, but also to evaluate some main characteristics of these models. This means, in other words, either formally proving that such properties are satisfied or guaranteeing that certain properties are not satisfied. And clearly, the builder then requires some automobile assistance to operate the device in such a manner that the property can be fulfilled. Two complementary learning methods to infer models from time series data have been explored in recent years: one is focused on the use of Solution Set Programming [Pau2011, Ben2017], the other on inductive logic programming [Rib2015, Rib2017]. We have been able to solve structures of hundreds of communicating components due to these approaches. However, both methods have the same downside, which is to have the unavoidable noise and/or inaccuracies in the results. For instance, where two similar results contribute to different actions at two different time stages, a crucial problem is to be able to explain that different outcomes arise from the same circumstances.

In order to resolve this limit, and if many models are consistent with some time data series, we plan to suggest a (semi)-automatic way of determining, from a modelling point of view, which kind of additional experiments should be carried out to choose the "real" one. This is important since some studies, owing to methodological limitations,

It can not be applied and any dynamical phase of a model can always not be observed. The purpose of the proposal is to contribute to a new computational paradigm focused on time series data to (partly) automate the simulation of complex systems. The resulting system will be followed by appropriate algorithms for learning and interpretation and the functional application of a free software tool.

Research subject & work plan: The method will be broken down into four key tasks: (1) collecting and curating data from DREAM Challenges (a common reverse engineering challenge) based on datasets given at the beginning of the work; (2) understanding the model (and its dynamics) based on the input of time series data. It is going to be important

Based on context information to

Determine the consistency of its components and cope with contradictory data; (3) Check the validity of a set of

Figure 1: Research plan to improve the learning process

Dynamic properties that help validate/invalidate the resulting model; (4) Algorithms designed to help I either propose new experiments to improve the uncertainty in the initial results, (ii) or change the model in such a way that the predicted behaviour may be satisfied (note that such adjustment might not be possible) if the expected properties are not met.

A tool published under a free-software licence with a GUI can execute the resulting algorithms.

References:

[Ben2017] E. Ben Abdallah, T. Ribeiro, M. Magnin, O. Roux and Katsumi Inoue. Modeling Delayed Dynamics in Biological Regulatory Networks from Time Series Data. Algorithms, Volume 10, Number 1, 2017. [Pau2011] L. Paulevé, M. Magnin, O. Roux. Tuning Temporal Features within the Stochastic π-Calculus. IEEE TSE, 37(6):858-871, 2011.

[Rib2015] T. Ribeiro, M. Magnin, K. Inoue, and C. Sakama. Learning Delayed Influences of Biological Systems. In Frontiers in Bioengineering and Biotechnology, 2, 81. 2015.

[Rib2017] T. Ribeiro, S. Tourret, M. Folschette, M. Magnin, D. Borzacchiello, P. Chinesta, O. F. Roux, K. Inoue. Inductive Learning from State Transitions over Continuous Domains. In International Conference on Inductive Logic Programming (pp. 124-139). Springer, 2017.