Evolution-Strategy Based, Fully Automatic, Numerical Optimization of Gas-Exchange Systems for IC Engines 2001-01-0577
Today, a number of simulation codes are available for pre-designing gas exchange systems of IC engines with good accuracy (e.g. PROMO, WAVE, GT-Power). However, optimizing such systems still requires numerous time consuming and inefficient trial and error runs. Also, accounting for constraints as size, volume, peak combustion pressure etc. multiplies the necessary efforts additionally. Hence there is a strong need for efficient procedures for finding optimum designs automatically and reliably.
To automatically find the global optimum design parameters under a given set of real constraints of a practical case, a multi-membered evolution-strategy based optimization code was developed. The code which efficiently finds the true optimum dimensions of gas exchange systems (duct lengths, duct diameters, volumes) of an IC engine. The code can be readily generalized, and adapted to arbitrary optimization problems.
The code is built around an evolution strategy based on 4 parents and 20 descendants. The properties of the descendants are statistically varied in each generation with a standard deviation of 0.2. As a test case a 3 valve single cylinder test engine was used with 10 free parameters of the gas exchange system. After a maximum of 3500 calls of the 1-D flow code (worst case), the global optimum is automatically found in all tests. Usually, optimum results are obtained for far less calls.
Cross checks of the performance of optimization strategies were carried out with gradient methods (fastest, if uni-modal) and 2 membered evolutionstrategies (for cases of low complexity). Whereas the gradient methods fail in multi-modal (i.e. normal) cases, the 2 membered evolution strategy is attractive only for low numbers of free parameters (e.g. ≤ 4). Experimental checks with conventional trial and error optimized engines showed that these engines could all be optimized further by implementing the global optimum design parameters predicted by our code.
In this paper the evolution-based optimization strategy is presented in detail and compared to other optimization strategies. The main features of the code are discussed and the results are shown for two typical cases.