669 research outputs found
Which Surrogate Works for Empirical Performance Modelling? A Case Study with Differential Evolution
It is not uncommon that meta-heuristic algorithms contain some intrinsic
parameters, the optimal configuration of which is crucial for achieving their
peak performance. However, evaluating the effectiveness of a configuration is
expensive, as it involves many costly runs of the target algorithm. Perhaps
surprisingly, it is possible to build a cheap-to-evaluate surrogate that models
the algorithm's empirical performance as a function of its parameters. Such
surrogates constitute an important building block for understanding algorithm
performance, algorithm portfolio/selection, and the automatic algorithm
configuration. In principle, many off-the-shelf machine learning techniques can
be used to build surrogates. In this paper, we take the differential evolution
(DE) as the baseline algorithm for proof-of-concept study. Regression models
are trained to model the DE's empirical performance given a parameter
configuration. In particular, we evaluate and compare four popular regression
algorithms both in terms of how well they predict the empirical performance
with respect to a particular parameter configuration, and also how well they
approximate the parameter versus the empirical performance landscapes
Evolutionary methods for modelling and control of linear and nonlinear systems
The aim of this work is to explore the potential and enhance the capability of evolutionary computation for the development of novel and advanced methodologies for engineering system modelling and controller design automation. The key to these modelling and design problems is optimisation.
Conventional calculus-based methods currently adopted in engineering optimisation are in essence local search techniques, which require derivative information and lack of robustness in solving practical engineering problems. One objective of this research is thus to develop an effective and reliable evolutionary algorithm for engineering applications. For this, a hybrid evolutionary algorithm is developed, which combines the global search power of a "generational" EA with the interactive local fine-tuning of Boltzmann learning. It overcomes the weakness in local exploration and chromosome stagnation usually encountered in pure EAs. A novel one-integer-one-parameter coding scheme is also developed to significantly reduce the quantisation error, chromosome length and processing overhead time. An "Elitist Direct Inheritance" technique is developed to incorporate with Bolzmann learning for reducing the control parameters and convergence time of EAs. Parallelism of the hybrid EA is also realised in this thesis with nearly linear pipelinability.
Generic model reduction and linearisation techniques in L2 and L∞ norms are developed based on the hybrid EA technique. They are applicable to both discrete and continuous-time systems in both the time and the frequency domains. Superior to conventional model reduction methods, the EA based techniques are capable of simultaneously recommending both an optimal order number and optimal parameters by a control gene used as a structural switch. This approach is extended to MIMO system linearisation from both a non-linear model and I/O data of the plant. It also allows linearisation for an entire operating region with the linear approximate-model network technique studied in this thesis.
To build an original model, evolutionary black-box and clear-box system identification
techniques are developed based on the L2 norm. These techniques can identify both the
system parameters and transport delay in the same evolution process. These open-loop
identification methods are further extended to closed-loop system identification. For robust
control, evolutionary L∞ identification techniques are developed. Since most practical
systems are nonlinear in nature and it is difficult to model the dominant dynamics of such a
system while retaining neglected dynamics for accuracy, evolutionary grey-box modelling
techniques are proposed. These techniques can utilise physical law dominated global clearbox
structure, with local black-boxes to include unmeasurable nonlinearities as the
coefficient models of the clear-box. This unveils a new way of engineering system
modelling.
With an accurately identified model, controller design problems still need to be overcome.
Design difficulties by conventional analytical and numerical means are discussed and a
design automation technique is then developed. This is again enabled by the hybrid
evolutionary algorithm in this thesis. More importantly, this technique enables the
unification of linear control system designs in both the time and the frequency domains
under performance satisfaction. It is also extended to control along a trajectory of operating
points for nonlinear systems. In addition, a multi-objective evolutionary algorithm is
developed to make the design more transparent and visible. To achieve a step towards
autonomy in building control systems, a technique for direct designs from plant step
response data is developed, which bypasses the system identification phase. These
computer-automated intelligent design methodologies are expected to offer added
productivity and quality of control systems
- …