Evolutionary Computation最新文献

Gathering sufficient instance data to either train algorithm-selection models or understand algorithm footprints within an instance space can be challenging. We propose an approach to generating synthetic instances that are tailored to perform well with respect to a target algorithm belonging to a predefined portfolio but are also diverse with respect to their features. Our approach uses a novelty search algorithm with a linearly weighted fitness function that balances novelty and performance to generate a large set of diverse and discriminatory instances in a single run of the algorithm. We consider two definitions of novelty: (1) with respect to discriminatory performance within a portfolio of solvers; (2) with respect to the features of the evolved instances. We evaluate the proposed method with respect to its ability to generate diverse and discriminatory instances in two domains (knapsack and bin-packing), comparing to another well-known quality diversity method, Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) and an evolutionary algorithm that only evolves for discriminatory behaviour. The results demonstrate that the novelty search method outperforms its competitors in terms of coverage of the space and its ability to generate instances that are diverse regarding the relative size of the "performance gap" between the target solver and the remaining solvers in the portfolio. Moreover, for the Knapsack domain, we also show that we are able to generate novel instances in regions of an instance space not covered by existing benchmarks using a portfolio of state-of-the-art solvers. Finally, we demonstrate that the method is robust to different portfolios of solvers (stochastic approaches, deterministic heuristics and state-of-the-art methods), thereby providing further evidence of its generality.

Evolutionary Computation (EC) often throws away learned knowledge as it is reset for each new problem addressed. Conversely, humans can learn from small-scale problems, retain this knowledge (plus functionality) and then successfully reuse them in larger-scale and/or related problems. Linking solutions to problems together has been achieved through layered learning, where an experimenter sets a series of simpler related problems to solve a more complex task. Recent works on Learning Classifier Systems (LCSs) has shown that knowledge reuse through the adoption of Code Fragments, GP-like tree-based programs, is plausible. However, random reuse is inefficient. Thus, the research question is how LCS can adopt a layered-learning framework, such that increasingly complex problems can be solved efficiently? An LCS (named XCSCF*) has been developed to include the required base axioms necessary for learning, refined methods for transfer learning and learning recast as a decomposition into a series of subordinate problems. These subordinate problems can be set as a curriculum by a teacher, but this does not mean that an agent can learn from it. Especially if it only extracts over-fitted knowledge of each problem rather than the underlying scalable patterns and functions. Results show that from a conventional tabula rasa, with only a vague notion of what subordinate problems might be relevant, XCSCF* captures the general logic behind the tested domains and therefore can solve any n-bit Multiplexer, n-bit Carry-one, n-bit Majority-on, and n-bit Even-parity problems. This work demonstrates a step towards continual learning as learned knowledge is effectively reused in subsequent problems.

We study the (1:s+1) success rule for controlling the population size of the (1,λ)- EA. It was shown by Hevia Fajardo and Sudholt that this parameter control mechanism can run into problems for large s if the fitness landscape is too easy. They conjectured that this problem is worst for the ONEMAX benchmark, since in some well-established sense ONEMAX is known to be the easiest fitness landscape. In this paper we disprove this conjecture. We show that there exist s and ɛ such that the self-adjusting (1,λ)-EA with the (1:s+1)-rule optimizes ONEMAX efficiently when started with ɛn zero-bits, but does not find the optimum in polynomial time on DYNAMIC BINVAL. Hence, we show that there are landscapes where the problem of the (1:s+1)-rule for controlling the population size of the (1,λ)-EA is more severe than for ONEMAX. The key insight is that, while ONEMAX is the easiest function for decreasing the distance to the optimum, it is not the easiest fitness landscape with respect to finding fitness-improving steps.

The fitness level method is a popular tool for analyzing the hitting time of elitist evolutionary algorithms. Its idea is to divide the search space into multiple fitness levels and estimate lower and upper bounds on the hitting time using transition probabilities between fitness levels. However, the lower bound generated by this method is often loose. An open question regarding the fitness level method is what are the tightest lower and upper time bounds that can be constructed based on transition probabilities between fitness levels. To answer this question, we combine drift analysis with fitness levels and define the tightest bound problem as a constrained multi-objective optimization problem subject to fitness levels. The tightest metric bounds by fitness levels are constructed and proven for the first time. Then linear bounds are derived from metric bounds and a framework is established that can be used to develop different fitness level methods for different types of linear bounds. The framework is generic and promising, as it can be used to draw tight time bounds on both fitness landscapes with and without shortcuts. This is demonstrated in the example of the (1+1) EA maximizing the TwoMax1 function.

Reproducibility of experiments is a complex task in stochastic methods such as evolutionary algorithms or metaheuristics in general. Many works from the literature give general guidelines to favor reproducibility. However, none of them provide both a practical set of steps or software tools to help in this process. In this article, we propose a practical methodology to favor reproducibility in optimization problems tackled with stochastic methods. This methodology is divided into three main steps, where the researcher is assisted by software tools which implement state-of-the-art techniques related to this process. The methodology has been applied to study the double-row facility layout problem (DRFLP) where we propose a new algorithm able to obtain better results than the state-of-the-art methods. To this aim, we have also replicated the previous methods in order to complete the study with a new set of larger instances. All the produced artifacts related to the methodology and the study of the target problem are available in Zenodo.

We argue that results produced by a heuristic optimisation algorithm cannot be considered reproducible unless the algorithm fully specifies what should be done with solutions generated outside the domain, even in the case of simple bound constraints. Currently, in the field of heuristic optimisation, such specification is rarely mentioned or investigated due to the assumed triviality or insignificance of this question. Here, we demonstrate that, at least in algorithms based on Differential Evolution, this choice induces notably different behaviours in terms of performance, disruptiveness, and population diversity. This is shown theoretically (where possible) for standard Differential Evolution in the absence of selection pressure and experimentally for the standard and state-of-the-art Differential Evolution variants, on a special test function and the BBOB benchmarking suite, respectively. Moreover, we demonstrate that the importance of this choice quickly grows with problem dimensionality. Differential Evolution is not at all special in this regard-there is no reason to presume that other heuristic optimisers are not equally affected by the aforementioned algorithmic choice. Thus, we urge the heuristic optimisation community to formalise and adopt the idea of a new algorithmic component in heuristic optimisers, which we refer to as the strategy of dealing with infeasible solutions. This component needs to be consistently: (a) specified in algorithmic descriptions to guarantee reproducibility of results, (b) studied to better understand its impact on an algorithm's performance in a wider sense (i.e., convergence time, robustness, etc.), and (c) included in the (automatic) design of algorithms. All of these should be done even for problems with bound constraints.

Reproducibility is important for having confidence in evolutionary machine learning algorithms. Although the focus of reproducibility is usually to recreate an aggregate prediction error score using fixed random seeds, this is not sufficient. Firstly, multiple runs of an algorithm, without a fixed random seed, should ideally return statistically equivalent results. Secondly, it should be confirmed whether the expected behaviour of an algorithm matches its actual behaviour, in terms of how an algorithm targets a reduction in prediction error. Confirming the behaviour of an algorithm is not possible when using a total error aggregate score. Using an error decomposition framework as a methodology for improving the reproducibility of results in evolutionary computation addresses both of these factors. By estimating decomposed error using multiple runs of an algorithm and multiple training sets, the framework provides a greater degree of certainty about the prediction error. Also, decomposing error into bias, variance due to the algorithm (internal variance), and variance due to the training data (external variance) more fully characterises evolutionary algorithms. This allows the behaviour of an algorithm to be confirmed. Applying the framework to a number of evolutionary algorithms shows that their expected behaviour can be different to their actual behaviour. Identifying a behaviour mismatch is important in terms of understanding how to further refine an algorithm as well as how to effectively apply an algorithm to a problem.

Describing the properties of complex systems that evolve over time is a crucial requirement for monitoring and understanding them. Signal Temporal Logic (STL) is a framework that proved to be effective for this aim because it is expressive and allows state properties as human-readable formulae. Crafting STL formulae that fit a particular system is, however, a difficult task. For this reason, a few approaches have been proposed recently for the automatic learning of STL formulae starting from observations of the system. In this paper, we propose BUSTLE (Bi-level Universal STL Evolver), an approach based on evolutionary computation for learning STL formulae from data. BUSTLE advances the state-of-the-art because it (i) applies to a broader class of problems, in terms of what is known about the state of the system during its observation, and (ii) generates both the structure and the values of the parameters of the formulae employing a bi-level search mechanism (global for the structure, local for the parameters). We consider two cases where (a) observations of the system in both anomalous and regular state are available, or (b) only observations of regular state are available. We experimentally evaluate BUSTLE on problem instances corresponding to the two cases and compare it against previous approaches. We show that the evolved STL formulae are effective and human-readable: the versatility of BUSTLE does not come at the cost of lower effectiveness.

Genetic Programming (GP) often uses large training sets and requires all individuals to be evaluated on all training cases during selection. Random down-sampled lexicase selection evaluates individuals on only a random subset of the training cases allowing for more individuals to be explored with the same amount of program executions. However, sampling randomly can exclude important cases from the down-sample for a number of generations, while cases that measure the same behavior (synonymous cases) may be overused. In this work, we introduce Informed Down-Sampled Lexicase Selection. This method leverages population statistics to build down-samples that contain more distinct and therefore informative training cases. Through an empirical investigation across two different GP systems (PushGP and Grammar-Guided GP), we find that informed down-sampling significantly outperforms random down-sampling on a set of contemporary program synthesis benchmark problems. Through an analysis of the created down-samples, we find that important training cases are included in the down-sample consistently across independent evolutionary runs and systems. We hypothesize that this improvement can be attributed to the ability of Informed Down-Sampled Lexicase Selection to maintain more specialist individuals over the course of evolution, while still benefiting from reduced per-evaluation costs.