Autonomous Decision Making Process Supporting Cognitive Waveform Design Nicolas Colson *† , Apostolos Kountouris * * France Telecom R&D, Orange Labs 28, Chemin du Vieux Chˆ ene, 38243 Meylan, France Email: nicolas.colson@orange-ftgroup.com Armelle Wautier † , Lionel Husson † † Department of Telecommunications, SUPELEC 91192 Gif-sur-Yvette, France Abstract—In this paper, we propose a cognitive decision making process driving the dynamic reconfiguration of a radio. The solution results from an original modeling of the cognitive design task based on the definition of two scales characterizing the solution space. It exploits the predictive capabilities of evolving connectionist systems improving their reliability through incremental learning as the radio interacts with its environment. The whole algorithm has been named RALFE for Reason And Learn From Experience based on its trial/error approach of the problem. This cognitive algorithm allows autonomous decision making with regard to multiple, possibly conflicting, operational objectives in a time-varying environment. The proposed approach is validated on a case of cognitive waveform design. I. I NTRODUCTION Cognitive radio [1], [2] is a technological concept pushing for the introduction of intelligent radio operation that goes beyond system adaptation and reconfiguration on the basis of simple criteria and rules. To become fully operational, a Cognitive Radio requires three main conditions. First, the device must be built on top of a Software Defined Radio platform [3] to benefit from its flexibility and agility. This condition makes the adaptation possible. Second, the de- vice must have a broad sensory activity in order to collect enough information to identify the context of the current communication: available services, radio link quality, service- dependent Quality of Service (QoS) requirements, regulation environment, remaining battery lifetime, etc. This condition is required to build a model of the context to adapt to. Third, the device must integrate a cognitive engine (CE) in order to solve, at runtime, the design problems formulated by the time- varying radio environment and the operational objectives set according to the context. This condition is central for cognitive behavior since it maps the context model to an appropriate action and thus corresponds to the intelligence per se. The cognitive behavior of a radio manifests itself through context awareness, multi-objective optimization capability, expertise based decision-making augmented by learning capabilities and last but not least transparent and computationally efficient operation. These are, in a nutshell, the high-level requirements of the decision making approach proposed in this paper. To our knowledge, the only comprehensive approach with similar scope is that of Virginia Tech [2, Chap. 7]. VTech’s CE makes use of advanced genetic algorithms techniques [4] to control the cognitive radio reconfiguration. A genetic al- gorithm is an efficient metaheuristic for solving optimization problems. Its success in optimization results from the parallel search of multiple solutions within the problem space. How- ever, it is often balanced by a slow convergence or a heavy computational load inherent to the evaluation of multiple solutions over several generations. The paper is organized as follows. Section II provides a general description of the proposed solution. Section III applies the solution to a simple case study of cognitive waveform design and presents experimental results validating the approach. Finally, Section IV concludes the paper and discusses some future work perspectives. II. THE PROPOSED SOLUTION A. Modeling of the cognitive design task Design objectives may be classified into two categories. Constrained objectives limit the search space by forbidding some alternatives violating one or several imposed constraints to meet (e.g. QoS requirements, regulation environment). Optimization objectives guide the decision making process toward an appropriate solution among the ones compatible with the constraints (e.g. download faster, save the battery energy). These two kinds of objectives are specified through the definition of bounding/optimization policies defined dy- namically in relation to the operational context evolution. This distinction between objectives lead us to the definition of two scales. The first scale, called the Performance Scale (PS), ranks the configurations according to their robustness toward the constraints. This scale is set from expert knowledge acquired beforehand through simulation and stored into the CE. For example, the robustness toward transmission errors may be deduced from the relative position of the Bit Error Rate (BER) curves produced for some predefined typical channel models. The second scale, called the Optimality Scale (OS), ranks the configurations according to their potential optimality with regard to the secondary objectives. This scale is set thanks to a grading system relying on configuration-specific performance indicators to assign grades according to the objectives impor- tance. For example, a configuration will be more satisfying in terms of throughput if it can transmit many information bits per symbol. However, attractive configurations are also more liable to violate the constraints due to their reduced robustness 978-1-4244-2644-7/08/$25.00 © 2008 IEEE