Search

The Multi-Prize Lottery Ticket Hypothesis posits that randomly initialized neural networks contain several subnetworks that achieve comparable accuracy to fully trained models of the same architecture. However, current methods require that the network is sufficiently overparameterized. In this work, we propose a modification to two state-of-the-art algorithms (Edge-Popup and Biprop) that finds high-accuracy subnetworks with no additional storage cost or scaling. The algorithm, Iterative Weight Recycling, identifies subsets of important weights within a randomly initialized network for intra-layer reuse. Empirically we show improvements on smaller network architectures and higher prune rates, finding that model sparsity can be increased through the "recycling" of existing weights. In addition to Iterative Weight Recycling, we complement the Multi-Prize Lottery Ticket Hypothesis with a reciprocal finding: high-accuracy, randomly initialized subnetwork's produce diverse masks, despite being generated with the same hyperparameter's and pruning strategy. We explore the landscapes of these masks, which show high variability.

Diversity is an important consideration in the construction of robust neural network ensembles. A collection of well trained models will generalize better if they are diverse in the patterns they respond to and the predictions they make. Diversity is especially important for low-cost ensemble methods because members often share network structure in order to avoid training several independent models from scratch. Diversity is traditionally analyzed by measuring differences between the outputs of models. However, this gives little insight into how knowledge representations differ between ensemble members. This paper introduces several interpretability methods that can be used to qualitatively analyze diversity. We demonstrate these techniques by comparing the diversity of feature representations between child networks using two low-cost ensemble algorithms, Snapshot Ensembles and Prune and Tune Ensembles. We use the same pre-trained parent network as a starting point for both methods which allows us to explore how feature representations evolve over time. This approach to diversity analysis can lead to valuable insights and new perspectives for how we measure and promote diversity in ensemble methods., 8 pages, 4 figures, Submitted to IJCNN-23

Rectified Linear Units (ReLU) are the default choice for activation functions in deep neural networks. While they demonstrate excellent empirical performance, ReLU activations can fall victim to the dead neuron problem. In these cases, the weights feeding into a neuron end up being pushed into a state where the neuron outputs zero for all inputs. Consequently, the gradient is also zero for all inputs, which means that the weights which feed into the neuron cannot update. The neuron is not able to recover from direct back propagation and model capacity is reduced as those parameters can no longer be further optimized. Inspired by a neurological process of the same name, we introduce Synaptic Stripping as a means to combat this dead neuron problem. By automatically removing problematic connections during training, we can regenerate dead neurons and significantly improve model capacity and parametric utilization. Synaptic Stripping is easy to implement and results in sparse networks that are more efficient than the dense networks they are derived from. We conduct several ablation studies to investigate these dynamics as a function of network width and depth and we conduct an exploration of Synaptic Stripping with Vision Transformers on a variety of benchmark datasets., Comment: 8 pages, 5 figures, Submitted to IJCNN-23

Neuroevolution is a promising area of research that combines evolutionary algorithms with neural networks. A popular subclass of neuroevolutionary methods, called evolution strategies, relies on dense noise perturbations to mutate networks, which can be sample inefficient and challenging for large models with millions of parameters. We introduce an approach to alleviating this problem by decomposing dense mutations into low-dimensional subspaces. Restricting mutations in this way can significantly reduce variance as networks can handle stronger perturbations while maintaining performance, which enables a more controlled and targeted evolution of deep networks. This approach is uniquely effective for the task of fine tuning pre-trained models, which is an increasingly valuable area of research as networks continue to scale in size and open source models become more widely available. Furthermore, we show how this work naturally connects to ensemble learning where sparse mutations encourage diversity among children such that their combined predictions can reliably improve performance. We conduct the first large scale exploration of neuroevolutionary fine tuning and ensembling on the notoriously difficult ImageNet dataset, where we see small generalization improvements with only a single evolutionary generation using nearly a dozen different deep neural network architectures., Comment: 8 Pages, 3 Figures

When working in an unfamiliar online environment, it can be helpful to have an observer that can intervene and guide a user toward a desirable outcome while avoiding undesirable outcomes or frustration. The Intervention Problem is deciding when to intervene in order to help a user. The Intervention Problem is similar to, but distinct from, Plan Recognition because the observer must not only recognize the intended goals of a user but also when to intervene to help the user when necessary. We formalize a family of Intervention Problems and show that how these problems can be solved using a combination of Plan Recognition methods and classification algorithms to decide whether to intervene. For our benchmarks, the classification algorithms dominate three recent Plan Recognition approaches. We then generalize these results to Human-Aware Intervention, where the observer must decide in real time whether to intervene human users solving a cognitively engaging puzzle. Using a revised feature set more appropriate to human behavior, we produce a learned model to recognize when a human user is about to trigger an undesirable outcome. We perform a human-subject study to evaluate the Human-Aware Intervention. We find that the revised model also dominates existing Plan Recognition algorithms in predicting Human-Aware Intervention.

Ensemble Learning is an effective method for improving generalization in machine learning. However, as state-of-the-art neural networks grow larger, the computational cost associated with training several independent networks becomes expensive. We introduce a fast, low-cost method for creating diverse ensembles of neural networks without needing to train multiple models from scratch. We do this by first training a single parent network. We then create child networks by cloning the parent and dramatically pruning the parameters of each child to create an ensemble of members with unique and diverse topologies. We then briefly train each child network for a small number of epochs, which now converge significantly faster when compared to training from scratch. We explore various ways to maximize diversity in the child networks, including the use of anti-random pruning and one-cycle tuning. This diversity enables "Prune and Tune" ensembles to achieve results that are competitive with traditional ensembles at a fraction of the training cost. We benchmark our approach against state of the art low-cost ensemble methods and display marked improvement in both accuracy and uncertainty estimation on CIFAR-10 and CIFAR-100., Comment: 12 pages, 7 figures, Accepted to AAAI-22; appendix added

Morphologically evolving robot systems need to include a learning period right after ‘birth’ to acquire a controller that fits the newly created body. In this paper, we investigate learning one skill in particular: walking in a given direction. To this end, we apply the HyperNEAT algorithm guided by a fitness function that balances the distance travelled in a direction and the deviation between the desired and the actually travelled directions. We validate this method on a variety of modular robots with different shapes and sizes and observe that the best controllers produce trajectories that accurately follow the correct direction and reach a considerable distance in the given test interval.

Discrete Mathematical Equations Governing HPA axis. Details the derivation of discrete equations employed for the simulation of HPA axis. (PDF 132 kb)

Asynchronous Simulation. Asynchronous update of the network contains two complex cyclic attractors. Since under asynchronous update each state might have more than two successors several state nodes in the graph present with multiple out-bound edges. (PDF 52 kb)

Priority Memory Update. Simulations in priority class with memory. CORT is placed in a second priority class update and the rest of the parameters in the model have a higher priority and frequency of update. (PDF 24 kb)

Stress Response Parameterization. Parameterization of stress response. (PDF 332 kb)

Interactive Parameter Tuning Algorithm. Describes the parameter identification algorithm and pseudo-code. (PDF 300 kb)

Synchronous Simulation. Simulations with synchronous update where each node has only one successor state. It is well known that synchronous update can result in spurious cycles [12]. (PDF 36 kb)

Performance of TSPmap with simulated datasets of different types of mapping populations (4way – four-way cross, bc – backcross, f2 – F2 population, riself – recombinant inbred lines), marker number (100, 400, 1000), proportions of missing data (0, 0.001, 0.01) and genotyping error rates (0, 0.001, 0.01). The accuracy of the TSPmap solution was measured by the correlation coefficient between the true marker order and marker order generated by TSPmap for each simulated dataset. Note the scale of the y-axis is 0.99935–1.000. (DOCX 40 kb)

Linkage maps generated by TSPmap using marker datasets from A. Arabidopsis thaliana [26] and B. & C. rice [27] compared to those generated by JoinMap. (DOCX 213Â kb)

Chicano, F., Whitley D., & Tinós R. (2016). Efficient Hill Climber for Multi-Objective Pseudo-Boolean Optimization. 16th European Conference on Evolutionary Computation for Combinatorial Optimization (LNCS 9595), pp. 88-103 Local search algorithms and iterated local search algorithms are a basic technique. Local search can be a stand-alone search method, but it can also be hybridized with evolutionary algorithms. Recently, it has been shown that it is possible to identify improving moves in Hamming neighborhoods for k-bounded pseudo-Boolean optimization problems in constant time. This means that local search does not need to enumerate neighborhoods to find improving moves. It also means that evolutionary algorithms do not need to use random mutation as a operator, except perhaps as a way to escape local optima. In this paper, we show how improving moves can be identified in constant time for multiobjective problems that are expressed as k-bounded pseudo-Boolean functions. In particular, multiobjective forms of NK Landscapes and Mk Landscapes are considered. Universidad de Málaga. Campus de Excelencia Internacional Andalucía Tech. Fulbright program, Ministerio de Educación (CAS12/00274), Ministerio de Economía y Competitividad (TIN2014-57341-R), Air Force Office of Scientific Research, Air Force Materiel Command, USAF (FA9550-11-1-0088), FAPESP (2015/06462-1) and CNPq.

Local search algorithms and iterated local search algorithms are a basic technique. Local search can be a stand along search methods, but it can also be hybridized with evolutionary algorithms. Recently, it has been shown that it is possible to identify improving moves in Hamming neighborhoods for k-bounded pseudo-Boolean optimization problems in constant time. This means that local search does not need to enumerate neighborhoods to find improving moves. It also means that evolutionary algorithms do not need to use random mutation as a operator, except perhaps as a way to escape local optima. In this paper, we show how improving moves can be identified in constant time for multiobjective problems that are expressed as k-bounded pseudo-Boolean functions. In particular, multiobjective forms of NK Landscapes and Mk Landscapes are considered., Comment: Paper accepted for publication in the 16th European Conference on Evolutionary Computation for Combinatorial Optimisation (EvoCOP 2016)

This paper introduces the NK Echo State Network. The problem of learning in the NK Echo State Network is reduced to the problem of optimizing a special form of a Spin Glass Problem known as an NK Landscape. No weight adjustment is used; all learning is accomplished by spinning up (turning on) or spinning down (turning off) neurons in order to find a combination of neurons that work together to achieve the desired computation. For special types of NK Landscapes, an exact global solution can be obtained in polynomial time using dynamic programming. The NK Echo State Network is applied to a reinforcement learning problem requiring a recurrent network: balancing two poles on a cart given no velocity information. Empirical results shows that the NK Echo State Network learns very rapidly and yields very good generalization.

There exist local search landscapes where the evaluation function is an eigenfunction of the graph Laplacian that corresponds to the neighborhood structure of the search space. Problems that dis- play this structure are called “Elementary Landscapes” and they have a number of special mathematical properties. The problems that are not elementary landscapes can be decomposed in a sum of elementary ones. This sum is called the elementary landscape decomposition of the problem. In this paper, we provide the elementary landscape decomposi- tion for the Hamiltonian Path Optimization Problem under two different neighborhoods. Universidad de Málaga. Campus de Excelencia Internacional Andalucía Tech.

The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) has been the most successful Evolution Strategy at exploiting covariance information; it uses a form of Principle Component Analysis which, under certain conditions, is suggested to converge to the correct covariance matrix, formulated as the inverse of the mathematically well-defined Hessian matrix. However, in practice, there exist conditions where CMA-ES converges to the global optimum (accomplishing its primary goal) while it does not learn the true covariance matrix (missing an auxiliary objective), likely due to step-size deficiency. These circumstances can involve high-dimensional landscapes with large condition numbers. This paper introduces a novel technique entitled Forced Optimal Covariance Adaptive Learning (FOCAL), with the explicit goal of determining the Hessian at the global basin of attraction. It begins by introducing theoretical foundations to the inverse relationship between the learned covariance and the Hessian matrices. FOCAL is then introduced and demonstrated to retrieve the Hessian matrix with high fidelity on both model landscapes and experimental Quantum Control systems, which are observed to possess a non-separable, non-quadratic search landscape. The recovered Hessian forms are corroborated by physical knowledge of the systems. This study constitutes an example for Natural Computing successfully serving other branches of natural sciences, and introducing at the same time a powerful generic method for any high-dimensional continuous search seeking landscape information.

International audience; Pareto local optimal solutions (PLOS) are believed to highly influence the dynamics and the performance of multi-objective optimization algorithms, especially those based on local search and Pareto dominance. A number of studies so far have investigated their impact on the difficulty of searching the landscape underlying a problem instance. However, the community still lacks knowledge on the structure of PLOS and the way it impacts the effectiveness of multi-objective algorithms. Inspired by the work on local optima networks in single-objective optimization, we introduce a PLOS network (PLOS-net) model as a step toward the fundamental understanding of multi-objective landscapes and search algorithms. Using a comprehensive set of ρmnk-landscapes, PLOS-nets are constructed by full enumeration, and selected network features are further extracted and analyzed with respect to instance characteristics. A correlation and regression analysis is then conducted to capture the importance of the PLOS-net features on the runtime and effectiveness of two prototypical Pareto-based heuristics. In particular, we are able to provide empirical evidence for the relevance of the PLOS-net model to explain algorithm performance. For instance, the degree of connectedness in the PLOS-net is shown to play an even more important role than the number of PLOS in the landscape.

We propose PHYSH (Parallel HYbridization for Simple Heuristics), a framework to ease the design and implementation of hybrid metaheuristics via cooperative parallelism. With this framework, the user only needs encode each of the desired metaheuristics and may rely on PHYSH for parallelization, cooperation and hybridization. PHYSH supports the combination of population-based and single-solution metaheuristics and enables the user to control the tradeoff between intensification and diversification. We also provide an open-source implementation of this framework which we use to model the Quadratic Assignment Problem (QAP) with a hybrid solver, combining three metaheuristics. We present experimental evidence that PHYSH brings significant improvements over competing approaches, as witness the performance on representative hard instances of QAP.

This paper proposes \(\text {Auto-MEKA}_{\text {GGP}}\), an Automated Machine Learning (Auto-ML) method for Multi-Label Classification (MLC) based on the MEKA tool, which offers a number of MLC algorithms. In MLC, each example can be associated with one or more class labels, making MLC problems harder than conventional (single-label) classification problems. Hence, it is essential to select an MLC algorithm and its configuration tailored (optimized) for the input dataset. \(\text {Auto-MEKA}_{\text {GGP}}\) addresses this problem with two key ideas. First, a large number of choices of MLC algorithms and configurations from MEKA are represented into a grammar. Second, our proposed Grammar-based Genetic Programming (GGP) method uses that grammar to search for the best MLC algorithm and configuration for the input dataset. \(\text {Auto-MEKA}_{\text {GGP}}\) was tested in 10 datasets and compared to two well-known MLC methods, namely Binary Relevance and Classifier Chain, and also compared to GA-Auto-MLC, a genetic algorithm we recently proposed for the same task. Two versions of \(\text {Auto-MEKA}_{\text {GGP}}\) were tested: a full version with the proposed grammar, and a simplified version where the grammar includes only the algorithmic components used by GA-Auto-MLC. Overall, the full version of \(\text {Auto-MEKA}_{\text {GGP}}\) achieved the best predictive accuracy among all five evaluated methods, being the winner in six out of the 10 datasets.

Given the availability of high-performing local search (LS) for single-objective (SO) optimisation problems, a successful approach to tackle their multi-objective (MO) counterparts is scalarisation-based local search (SBLS). SBLS strategies solve multiple scalarisations, aggregations of the multiple objectives into a single scalar value, with varying weights. They have been shown to work specially well as the initialisation phase of other types of MO local search, e.g., Pareto local search (PLS). A drawback of existing SBLS strategies is that the underlying SO-LS method is unaware of the MO nature of the problem and returns only a single solution, discarding any intermediate solutions that may be of interest. We propose here two new SBLS initialisation strategies (ChangeRestart and ChangeDirection) that overcome this drawback by augmenting the underlying SO-LS method with an archive of nondominated solutions used to dynamically update the scalarisations. The new strategies produce better results on the bi-objective permutation flowshop problem than other five SBLS strategies from the literature, not only on their own but also when used as the initialisation phase of PLS.

For multiobjective optimization problems with uncertain parameters in the objective functions, different variants of minmax robustness concepts have been defined in the literature. The idea of minmax robustness is to optimize in the worst case such that the solutions have the best objective function values even when the worst case happens. However, the computation of the minmax robust Pareto optimal solutions remains challenging. This paper proposes a simple indicator based evolutionary algorithm for robustness (SIBEA-R) to address this challenge by computing a set of non-dominated set-based minmax robust solutions. In SIBEA-R, we consider the set of objective function values in the worst case of each solution. We propose a set-based non-dominated sorting to compare the objective function values using the definition of lower set less order for set-based dominance. We illustrate the usage of SIBEA-R with two example problems. In addition, utilization of the computed set of solutions with SIBEA-R for decision making is also demonstrated. The SIBEA-R method shows significant promise for finding non-dominated set-based minmax robust solutions. peerReviewed

Over the years, many interactive multiobjective optimization methods based on a reference point have been proposed. With a reference point, the decision maker indicates desirable objective function values to iteratively direct the solution process. However, when analyzing the performance of these methods, a critical issue is how to systematically involve decision makers. A recent approach to this problem is to replace a decision maker with an artificial one to be able to systematically evaluate and compare reference point based interactive methods in controlled experiments. In this study, a new artificial decision maker is proposed, which reuses the dynamics of particle swarm optimization for guiding the generation of consecutive reference points, hence, replacing the decision maker in preference articulation. We use the artificial decision maker to compare interactive methods. We demonstrate the artificial decision maker using the DTLZ benchmark problems with 3, 5 and 7 objectives to compare R-NSGA-II and WASF-GA as interactive methods. The experimental results show that the proposed artificial decision maker is useful and efficient. It offers an intuitive and flexible mechanism to capture the current context when testing interactive methods for decision making. Ministerio de Ciencia, Innovación y Universidades TIN2017-86049-R Junta de Andalucía P12-TIC-1519

In this work we present a comparison of different adaptive methods for the flexible job shop scheduling problem with due-dates, release-dates and machine perturbations. We look at online and offline scheduling approaches with different known information. We generated several new problem instances using classic flexible job shop scheduling instances as a base problem. Using these instances we tested and compared the approaches introduced here.