Performance prediction has been a key part of the neural architecture search (NAS) process, allowing to speed up NAS algorithms by avoiding resource-consuming network training. Although many performance predictors correlate well with ground truth performance, they require training data in the form of trained networks. Recently, zero-cost proxies have been proposed as an efficient method to estimate network performance without any training. However, they are still poorly understood, exhibit biases with network properties, and their performance is limited. Inspired by the drawbacks of zero-cost proxies, we propose neural graph features (GRAF), simple to compute properties of architectural graphs. GRAF offers fast and interpretable performance prediction while outperforming zero-cost proxies and other common encodings. In combination with other zero-cost proxies, GRAF outperforms most existing performance predictors at a fraction of the cost.

Cooperation with previously unseen agents in multi-agent cooperative scenario is a challenging problem. Agents trained together tend to rely on certain conventions discovered during the training that may not be present when they should cooperate with new agents. This leads to unsatisfying performance in such cases. We call this the off-diagonal problem. In this paper, we investigate this problem on more than 20 maps from the Overcooked environment. First, we propose and evaluate a number of metrics to quantify the off-diagonal problem and then we propose a population-based training technique to alleviate this problem. The results show that the proposed metrics can be used to divide the maps into groups based on how much they are affected by this problem, and the population-based training improves the performance of the agents especially in those maps, where the problem is present, without having negative consequences in the other maps.

In dynamic environments, external events might occur and modify the environment without consent of intelligent agents. Plans of the agents might hence be disrupted and, worse, the agents might end up in dead-end states and no longer be able to achieve their goals. Hence, the agents should monitor the environment during plan execution and if they encounter a dangerous situation they should (reactively) act to escape from it.In this paper, we introduce the notion of dangerous states that the agent might encounter during its plan execution in dynamic environments. We present a method for computing lower bound of dangerousness of a state after applying a sequence of actions. That method is leveraged in identifying situations in which the agent has to start acting to avoid danger. We present two types of such behaviour – purely reactive and proactive (eliminating the source of danger). The introduced concepts for planning with dangerous states are implemented and tested in two scenarios – a simple RPG-like game, called Dark Dungeon, and a platform game inspired by the Perestroika video game. The results show that reasoning with dangerous states achieves better success rate (reaching the goals) than naive planning or rule-based techniques.

The problem of coordinating the charging of electric vehicles gains more importance as the number of such vehicles grows. In this paper, we develop a method for the training of controllers for the coordination of EV charging. In contrast to most existing works on this topic, we require the controllers to preserve the privacy of the users, therefore we do not allow any communication from the controller to any third party.In order to train the controllers, we use the idea of imitation learning – we first find an optimum solution for a relaxed version of the problem using quadratic optimization and then train the controllers to imitate this solution. We also investigate the effects of regularization of the optimum solution on the performance of the controllers. The method is evaluated on realistic data and shows improved performance and training speed compared to similar controllers trained using evolutionary algorithms.

Deep learning has achieved great results in the last decade, however, it is sensitive to so called adversarial attacks - small perturbations of the input that cause the network to classify incorrectly. In the last years a number of attacks and defenses against these attacks were described. Most of the defenses however focus on defending against gradient-based attacks. In this paper, we describe an evolutionary attack and show that the adversarial examples produced by the attack have different features than those from gradient-based attacks. We also show that these features mean that one of the state-of-the-art defenses fails to detect such attacks.

Planning and acting in dynamic environments deals with non-deterministic events that might change the state of the environment without consent of the agent. In the worst case, some events might cause the agent to become ``trapped'' in a dead-end state, which in practice might mean damage or destruction of the agent. Presence of non-deterministic events often considerably increases the number of alternatives that might occur in a single step and hence traditional non-deterministic planning techniques might not scale. In this paper, we address a class of problems where non-deterministic events represent ``cyclic phenomena''. If they interfere with the agent, they might be dangerous for it (e.g. ships cruising through the area of AUV operations). We present techniques that initially analyse the problem whether it falls within this class by considering the notion of event reversibility and if so, these techniques generate a plan such that encountered unsafe states, in which the ``cyclic phenomena'' might interfere with the agent, can be eventually crossed without any risk of ``falling'' into a dead-end state. Our approach is evaluated in the AUV and Perestroika domains.

Deep neural networks are known to be sensitive to adversarial examples - inputs that are created in such a way that they are similar (if viewed by people) to clean inputs, but the neural network has high confidence that they belong to another class. In this paper, we study a new type of neural network unit similar to the convolutional units, but with a more local behavior. The unit is based on the Gaussian radial basis function. We show that if we replace the first convolutional layer in a convolutional network by the new layer (called RBFolutional), we obtain better robustness towards adversarial samples on the MNIST and CIFAR10 datasets, without sacrificing the performance on the clean examples.

We propose a generative model for artificial composition of both classical and popular music with the goal of producing music as well as humans do. The problem is that music is based on a highly sophisticated hierarchical structure and it is hard to measure its quality automatically. Contrary to other's work, we try to generate a symbolic representation of music with multiple different instruments playing simultaneously to cover a broader musical space. We train three modules based on LSTM networks to generate the music; a lot of effort is put into reducing the high complexity of multi-instrumental music representation by a thorough musical analysis. We believe that the proposed preprocessing techniques and symbolic representation constitute a useful resource for future research in this field.

We present an online parallel portfolio selection algorithm based on the island model, which is commonly used for parallelization of evolutionary algorithms. In our case each of the islands runs a different optimization algorithm. The distributed computation is managed by a central planner which periodically changes the running methods during the execution of the algorithm - less successful methods are removed while new instances of more successful methods are added. We compare different types of planners in the heterogeneous island model among themselves and also to the traditional homogeneous model on a wide set of problems. The tests include experiments with different representations of the individuals and different duration of fitness function evaluations. The results show that heterogeneous models are a more general and universal computational tool compared to homogeneous models.

We describe first results on a traffic-lights controller based on neural networks optimized by an evolutionary algorithm. Among the inputs of the neural network are outputs of other popular control algorithms, thus the evolved controller can be considered and ensemble controller. In a series of experiments, we show that evolution is capable of creating controllers that provide promising performance better than any of a number of baselines.

We propose and evaluate controllers for the coordination of the charging of electric vehicles. The controllers are based on neural networks and are completely de-centralized, in the sense that the charging current is completely decided by the controller itself. One of the versions of the controllers does not require any outside communication at all. We test controllers based on two different architectures of neural networks-the feed-forward networks and the echo state networks. The networks are optimized by either an evolutionary algorithm (CMA-ES) or by a gradient-based method. The results of the different architectures and the different optimization algorithms are compared in a realistic scenario. We show that the controllers are able to charge the cars while keeping the peak consumptions almost the same as when no charging is performed. Moreover, the controllers fill the valleys of the consumption thus reducing the difference between the maximum and minimum consumption in the grid.

We present a parallel evolutionary algorithm with interleaving generations. The algorithm uses a careful analysis of genetic operators and selection in order to evaluate individuals from following generations while the current generation is still not completely evaluated. This brings significant advantages in cases where each fitness evaluation takes different amount of time, the evaluations are performed in parallel, and a traditional generational evolutionary algorithm has to wait for all evaluations to finish. The proposed algorithm provides better utilization of computational resources in these cases. Moreover, the algorithm is functionally equivalent to the generational evolutionary algorithm, and thus it does not have any evaluation time bias, which is often present in asynchronous evolutionary algorithms.

Acting of intelligent entities, or agents, is determined by their ability to deliberatively reason towards longer-term goals while being able to react to unexpected situations. In particular, the agent generates a plan and while the plan is being executed the agent monitors the environment and eventually reacts to avoid possibly dangerous situations. In this paper, we present a concept of "safe'' planning and execution framework that defines the notion of "dangerous'' states the agent should avoid and evaluates how agent's possible actions influence "dangerousness'' of its situation. In particular, we introduce three "clever'' agents accommodating such a framework providing three different strategies of handling "dangerous'' situations occurring during the plan execution.

In decomposition based algorithms the quality of the resulting solutions depends on the weights used in the decomposition scheme. Usually the weights are generated in the beginning and remain fixed during the evolution, which may lead to poor distribution of solutions along the Pareto front. In this paper, we describe an extension of the popular MOEA/D algorithm which is able to tune the weights in order to find a set of solutions which maximizes a user specified objective. This adaptation is added as a new step to the algorithm which uses an approximation of the Pareto front to find suitable points in the objective space. These points are translated back into weights in such way to lead MOEA/D to find these points.

This paper describes an algorithm for the automated design of whole machine learning workflows, including preprocessing of the data and automatic creation of several types of ensembles. The algorithm is based on strongly typed genetic programming which ensures the validity of the workflows. The evolution of the individuals in the population is asynchronous in order to improve the utilization of computational resources. The approach is validated on four data sets from the UCI machine learning repository.

Metalearning deals with an important problem in machine-learning, namely selecting the right techniques to model the data at hand. In most of the metalearning approaches, a notion of similarity between datasets is needed. Our approach derives the similarity measure by combining arbitrary attribute similarity functions ordered by the optimal attribute assignment. In this paper, we propose a genetic programming based approach to the evolution of an attribute similarity inducing function. The function is composed of two parts - one describes the similarity of categorical attributes, the other describes the similarity of numerical attributes. Co-evolution is used to put these two parts together to form the similarity function. We use a repairing approach to guarantee some of the metric features for this function, and also discuss which of these features are important in metalearning.

The resulting set of solutions obtained by MOEA/D depends on the weights used in the decomposition. In this work, we use this feature to incorporate user preferences into the search. We use co-evolutionary approach to change the weights adaptively during the run of the algorithm. After the user specifies their preferences by assigning binary preference values to the individuals, the co-evolutionary step improves the distribution of weights by creating new (offspring) weights and selecting those that better match the user preferences. The algorithm is tested on a set of benchmark functions with a set of different user preferences.

Evolutionary algorithms are among the best multi-objective optimizers, but the large number of objective function evaluations they require makes it hard to use them to solve certain real-life tasks. In this work we present a surrogate-based local search for the multi-objective covariance matrix adaption evolution strategy (MO-CMA-ES). The local search is based on the estimation of hypervolume contribution of each individual and maximization of this contribution. This work extends our previous work and makes such surrogate models applicable to problems with more than two objectives. Moreover, it uses a unique feature of MO-CMA-ES to make the local search more effective. The results indicate that the algorithm can find solutions of the same quality as MO-CMA-ES while using 30-50 percent less objective function evaluations.

Several contemporary multi-objective surrogate-based algorithms use some kind of local search operator. The search technique used in this operator can largely affect the performance of the multi-objective optimizer as a whole, however, little attention is often paid to the selection of this technique. In this paper, we compare three different local search techniques and evaluate their effect on the performance of two different surrogate based multi-objective optimizers. The algorithms are evaluated using the well known ZDT and WFG benchmark suites and recommendations are made based on the results.

This paper describes a surrogate based multi-objective evolutionary algorithm with hyper-volume contribution-based local search. The algorithm switches between an NSGA-II phase and a local search phase. In the local search phase, a model for each of the objectives is trained and CMA-ES is used to optimize the hyper-volume contribution of each individual with respect to its two neighbors on the non-dominated front. The performance of the algorithm is evaluated using the well known ZDT and WFG benchmark suites.

In surrogate evolutionary algorithms, usually the type of surrogate model is chosen beforehand, and it is never changed during the run of the evolution. Moreover, the reasoning why a particular type of model was chosen is often missing. In this paper, we present a framework which in each generation selects the most suitable surrogate from a set of models based on some pre-defined criteria. The results based on different types of model selectors are compared, and the dynamics of the evolution together with the change of the selected model type during the run of the evolutionary algorithm are discussed.

We define a new class of 2D-stock cutting problems, the so called semi-quillotinable problems, and show its practical importance in solving the task of creating optimal cutting plans for a circular saw. Furthermore, we create a new algorithm suited for solving of semi-guillotinable problems by adapting existing evolutionary algorithms for both guillotinable and non-guillotinable 2D stock cutting problems. This algorithm is compared to standard algorithms on a selected set of both benchmark and real-life problems.

The paper presents a surrogate-based evolutionary strategy for multiobjective optimization. The evolutionary strategy uses distance based aggregate surrogate models in two ways: as a part of memetic search and as way to pre-select individuals in order to avoid evaluation of bad individuals. The model predicts the distance of individuals to the currently known Pareto set. The newly proposed algorithm is compared to other algorithms which use similar surrogate models on a set of benchmark functions.

The goal of our data-mining multi-agent system is to facilitate data-mining experiments without the necessary knowledge of the most suitable machine learning method and its parameters to the data. In order to replace the expertâs knowledge, the meta-learning subsystems are proposed including the parameter-space search and method recommendation based on previous experiments. In this paper we show the results of the parameter-space search with several search algorithms â" tabulation, random search, simmulated annealing, and genetic algorithm.

Most existing surrogate based evolutionary algorithms deal with only one model selected by the authors and different models are not considered. In this paper we propose a framework which enables automatic selection of types of surrogate models, and evaluate the effect of the type of selection on the overall performance of the resulting evolutionary algorithm. Two different types of model selection are tested and compared both in pre-selection scenario and in local search scenario.

In this paper we present a multiobjective evolutionary algorithm which uses surrogate models in two different ways -- during a local search and during pre-selection. Two different approaches to surrogate modeling are used, and the algorithm provides multiple individuals in each generation to enable easy parallelization. The algorithm is tested and compared to standard multiobjective evolutionary algorithms and to our previously developed surrogate evolution strategy. We also discuss the importance of the use of two different approaches and show that it improves the convergence speed significantly.

Evolutionary algorithms generally require a large number of objective function evaluations which can be costly in practice. These evaluations can be replaced by evaluations of a cheaper meta-model (surrogate model) of the objective functions. In this paper we present a novel distance based aggregate surrogate model for multiobjective optimization and describe a memetic multiobjective algorithm based on this model. Various variants of the models are tested and discussed and the algorithm is compared to standard multiobjective evolutionary algorithms. We show that our algorithm greatly reduces the number of required objective function evaluations.

In the field of multi-objective optimization there have been attempts to reduce the number of objective function evaluations by the use of surrogate models. However, in many-objective optimization, this work still has to be done to make the optimizers more practically usable. In this paper we show, that aggregate meta-models can be used even for the many-objective optimization and that they can also improve the performance of the many-objective optimizer. Moreover, meta-models are discussed from another point of view and compared to scalarization techniques in many-objective optimization. Two algorithms using our models are compared to IBEA on a set of selected benchmark functions with 5, 10, and 15 objectives.

The majority of multiobjective genetic algorithms is computationally expensive, therefore they often need to be parallelized before they can be used to solve practical tasks. Parallelization of multiobjective genetic algorithms is a relatively studied area, but no clearly winning approach has appeared yet. In this paper we present a novel parallel hybrid algorithm which combines multiobjective and single-objective genetic algorithms. We show that this algorithm can be successfully used to solve multiobjective optimization problems while outperforming more traditional parallel versions of multiobjective genetic algorithms.