Browsing by Subject "Evolutionary algorithms"

Now showing 1 - 3 of 3

Accelerating evolution through gene masking and distributed search
(2023-04-17) Shahrzad, Hormoz; Miikkulainen, Risto
In building practical applications of evolutionary computation (EC), two optimizations are essential. First, the parameters of the search method need to be tuned to the domain in order to balance exploration and exploitation effectively. Second, the search method needs to be distributed to take advantage of parallel computing resources. This paper presents BLADE (BLAnket Distributed Evolution) as an approach to achieving both goals simultaneously. BLADE uses blankets (i.e., masks on the genetic representation) to tune the evolutionary operators during the search, and implements the search through hub-and-spoke distribution. In the thesis, (1) the blanket method is formalized for the (1 + 1)EA case as a Markov chain process. Its effectiveness is then demonstrated by analyzing dominant and subdominant eigenvalues of stochastic matrices, suggesting a generalizable theory; (2) the fitness-level theory is used to analyze the distribution method; and (3) these insights are verified experimentally on three benchmark problems, showing that both blankets and distribution lead to accelerated evolution. Moreover, a surprising synergy emerges between them: When combined with distribution, the blanket approach achieves more than n-fold speedup with n clients in some cases. The work thus highlights the importance and potential of optimizing evolutionary computation in practical applications.
Evolutionary algorithms in optimization of technical rules for automated stock trading
(2004-12-18) Subramanian, Harish K.; Stone, Peter, 1971-; Kuipers, Benjamin
The effectiveness of technical analysis indicators as a means of predicting future price levels and enhancing trading profitability in stock markets is an issue constantly under review. It is an area that has been researched and its profitability examined in foreign exchange trade [1], portfolio management [2] and day trading [3]. Their use has been advocated by many traders [4], [5] and the uses of these charting and analysis techniques are being scrutinized [6], [7]. However, despite their popularity among human traders, a number of popular technical trading rules can be loss-making when applied individually, typically because human technical traders use combinations [8], [9] of a broad range of these technical indicators. Moreover, successful traders tend to adapt to market conditions by varying the weight they give to certain trading rules and dropping some of them as they are deemed to be loss-making. In this thesis, we try to emulate such a strategy by developing trading systems consisting of rules based on combinations of different indicators, and evaluating their profitability in a simulated economy. We propose and empirically examine two schemes, using evolutionary algorithms (genetic algorithm and genetic programming), of optimizing the combination of technical rules. A multiple model approach [10a] is used to control agent behavior and encourage unwinding of share position to ensure a zero final share position (as is essential within the framework that our experiments are run in). Evaluation of the evolutionary composite technical trading strategies leads us to believe that there is substantial merit in such evolutionary designs (particularly the weighted majority model), provided the right learning parameters are used. To explore this possibility, we evaluated a fitness function measure limiting only downside volatility, and compared its behavior and benefits with the classical Sharpe ratio, which uses a measure of standard deviation. The improved performance of the new fitness function strengthens our claim that a weighted majority approach could indeed be useful, albeit with a more sophisticated fitness function
Evolutionary controllers for identifying viable regimes and obtaining optimal performance in precision inkjet systems
(2019-12-06) Snyder, Brent Andrew; Sreenivasan, S.V.; Djurdjanovic, Dragan; Longoria, Raul G; LaBrake, Dwayne
Drop on demand piezoelectric inkjets have become an important device for direct patterning and adaptive material deposition in micro- and nano-fabrication applications. Key performance, reliability, and cost metrics for piezo-jets include drop volume minimization, drop volume accuracy and precision, drop placement accuracy, drop velocity, hours of continuous operation, and throughput. Accurately predicting drop formation from an actuation waveform using physics-based models is challenging as such models require knowledge of several inkjet parameters that cannot be determined non-destructively. Current practice involves ad-hoc manual recalibration of the actuation waveform to obtain reliable jetting of a variety of materials. This prevents the use of higher order waveforms defined by large numbers of parameters, where such waveforms have been demonstrated in the literature to achieve more aggressive performance metrics than lower order waveforms. This dissertation presents automatic piezo-jet waveform tuning methods based on evolutionary algorithms and computer-vision-based monitoring drops in-flight and drops that have been dispensed onto a substrate. Actuating and monitoring a piezo-jet as part of a “machine-in-the-loop” optimization scheme circumvents the need for complex forward models, as key performance metrics are estimated from images of actual jetted drops. Automatic tuning also enables exploration of previously unachievable highly complicated higher order waveforms comprised of more than a hundred parameters. In this dissertation, three fixed waveform topologies of increasing complexity were applied to optimize waveforms using genetic algorithms (GA) for a single-nozzle inkjet based on computer vision feedback from in-flight drop monitoring. These GA experiments automatically found waveforms for water and ethyl acetate, wherein the latter is considered rheologically impossible to jet based on fluid mechanics studies in the literature, but which was jetted at drop volumes measuring 0.8 pL continuously for several hours without faults. The resulting ratio between the 11.5 µm drop diameter and the 50 µm inkjet nozzle aperture was an impressive 23%. Next, a novel optimization scheme of a GA with a variable-length or “free” topology was developed to optimize waveforms for a more complex multi-nozzle piezo-jet using top-down imaging of drops dispensed on a silicon wafer to measure performance. This free-topology GA enabled exploration of highly sophisticated controllers and resulted in waveforms with as many as 124 parameters that reduced drop volume by 22.8% to an estimated 336 fL as compared to the waveforms found by a fixed-topology GA with 13 parameters.