Ramin Mehran's homepage


Summery of Master’s Thesis

Dynamic Modeling of Context in Control

by: Ramin Mehran


In this thesis, the role of context in control design and system modeling has been surveyed and a new controller architecture based on the function of context in human mind has been proposed. In addition, based on previous works on system modeling, two enhanced modeling algorithms, which include the properties of context-based systems are proposed and compared. The proposed control architecture is used to solve the cart and pole problem without prior knowledge about system dynamic through online learning a neuro-fuzzy controller.


Namely, the context in intelligent systems classified into two main categories: Static and Dynamic context. In this research, the context is defined as any sub-space states that user can assign different mood or phases of the system. From that angle, every system that classifies state variables in a subspace and acts differently for each case is viewed as a context-based system. Therefore, a control system based on static context is define as a system with fixed clusters in a subset of the system states, which has controlling laws for them. On the other hand, a system based on dynamic context is a system that learns to cluster context states and learns the controlling law for them. Moreover, in dynamic context, a system should also forget old context and learn to create new cases. Humans are able to use dynamic contexts in language and their actions. This is the genesis of the inspirations of this thesis to create hierarchical architecture with online clustering for control systems based on cognitive models [1, 2].


In the first section of the thesis, the algorithm of Locally Linear Model Trees for modeling of systems is considered as a system with static context. In this algorithm, the partitioning of the input state is based on a heuristic method. In this thesis, this method is evaluated, an enhancing partitioning based on particle swarm is introduced, and results are compared. 


In the second section, to pursue the design of control systems with dynamic context, a new modeling architecture has introduced improving the previous works of Kasabov [3] on dynamic evolving neuro-fuzzy system. In this research, this algorithm is enhanced in both performance and robustness with the ideas inspired from swarm foraging. Every cluster is considered as a swarm that has a local memory and some updating rule for center and radius changing. The new algorithm comprises the following enhancements over [3]:

  • Robustness to noise
  • Ability to shrink the clusters: Necessary to forget old clusters
  • Better clustering performance by using personal memory of best position for each cluster


This new algorithm is also the basis for the controller architecture that acts on the dynamic context.


The main contribution of this thesis is the design of a hierarchical adaptive control architecture, which works on a dynamic context. The controller perceives context in the state space and weights the lower level controllers according to it. The lower controllers are designed separately without knowledge about other controllers. Therefore, in this architecture, a hierarchical control design with low complexity can be used to solve complex control problems. In this architecture, the detection of context subspaces is done by online clustering of the state space with help of a learning method similar to the previously mentioned modeling method. Simultaneously, this is accompanied by adapting of locally linear models that construct the weighting system with weighted RLS. Therefore, both the structure and the parameters of the controller are learned through experience.

Figure 1. The proposed controller architecture


[1] M.A. Arbib, Language Evolution: The Mirror System Hypothesis, M.A. Arbib (ed.), Handbook of Brain Theory and Neural Networks, MIT Press, 2002.


[2] Christian Balkenius, Jan Morén A Computational Model of Context Processing, Proceedings of the 6th International Conference on the Simulation of Adaptive Behaviour, Cambridge, Mass., 2000. The MIT Press


[3] Nikola K. Kasabov, Qun Song, DENFIS: Dynamic Evolving Neural-Fuzzy Inference System and Its Application for Time-Series Prediction, IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 10, NO. 2, APRIL 2002 


- The final presentation of the project: Dynamic Modeling of Context in Control [pdf]

- The introductory presentation of my project is available now:

"Dynamic Context in Control - Introductory Seminar" [pdf]

- The second seminar on: Behavior Based Systems is available now: seminar2.pdf

- My final presentation for seminar course: Behavior Based Systems - Literature Review [PPT]

^ back to top ^

Last update: 31-12-06 Ó Copyrighted by Ramin Mehran 2004.