Ing. Jan Drchal, Ph.D.

Two challenges can be found in a life-long classifier that learns continually: the concept drift, when the probability distribution of data is changing in time, and catastrophic forgetting when the earlier learned knowledge is lost. There are many proposed solutions to each challenge, but very little research is done to solve both challenges simultaneously. We show that both the concept drift and catastrophic forgetting are closely related to our proposed description of the life-long continual classification. We describe the process of continual learning as a wrap modification, where a wrap is a manifold that can be trained to cover or uncover a given set of samples. The notion of wraps and their cover/uncover modifiers are theoretical building blocks of a novel general life-long learning scheme, implemented as an ensemble of variational autoencoders. The proposed algorithm is examined on evaluation scenarios for continual learning and compared to state-of-the-art algorithms demonstrating the robustness to catastrophic forgetting and adaptability to concept drift but also showing the new challenges of the life-long classification.

Dubins tours represent a solution of the Dubins traveling salesman problem (DTSP) that is a variant of the optimization routing problem to determine a curvature-constrained shortest path to visit a set of locations such that the path is feasible for Dubins vehicle, which moves only forward and has a limited turning radius. The DTSP combines the NP-hard combinatorial optimization to determine the optimal sequence of visits to the locations, as in the regular TSP, with the continuous optimization of the heading angles at the locations, where the optimal heading values depend on the sequence of visits and vice versa. We address the computationally challenging DTSP by fast evaluation of the sequence of visits by the proposed windowing surrogate model (WiSM), which estimates the length of the optimal Dubins path connecting a sequence of locations in a Dubins tour. The estimation is sped up by a regression model trained using close to optimum solutions of small Dubins tours that are generalized for large-scale instances of the addressed DTSP utilizing the sliding-window technique and a cache for already computed results. The reported results support that the proposed WiSM enables fast convergence of a relatively simple evolutionary algorithm to high-quality solutions of the DTSP. We show that with an increasing number of locations, our algorithm scales significantly better than other state-of-the-art DTSP solvers.

Multi-goal curvature-constrained planning such as the Dubins Traveling Salesman Problem (DTSP) combines NP-hard combinatorial routing with continuous optimization to determine the optimal vehicle heading angle for each target location. The problem can be addressed as combinatorial routing using a finite set of heading samples at target locations. In such a case, optimal heading samples can be determined for a sequence of targets in polynomial time, and the DTSP can be solved as searching for a sequence with the minimal cost. However, the examination of sequences can be computationally demanding for high numbers of heading samples and target locations. A fast rejection schema is proposed to quickly examine unfavorable sequences using lower bound estimation of Dubins tour cost based on windowing technique that evaluates short subtours of the sequences. Furthermore, the computation using small problem instances can benefit from reusing stored results and thus speed up the search. The reported results indicate that the computational burden is decreased about two orders of magnitude, and the proposed approach supports finding high-quality solutions of routing problems with Dubins vehicle.

A life-long classifier that learns incrementally has many challenges such as concept drift, when the class changes in time, and catastrophic forgetting when the earlier learned knowledge is lost. Many successful connectionist solutions are based on an idea that new data are learned only in a part of a network that is relevant to the new data. We leverage this idea and propose a novel method for learning an ensemble of specialized autoencoders. We interpret autoencoders as manifolds that can be trained to contain or exclude given points from the input space. This manifold manipulation allows us to implement a classifier that can suppress catastrophic forgetting and adapt to concept drift. The proposed algorithm is evaluated on an incremental version of the XOR problem and on an incremental version of the MNIST classification where we achieved 0.9 accuracy which is a significant improvement over the previously published results

Evaluation of incremental classification algorithms is a complex task because there are many aspects to evaluate. Besides the aspects such as accuracy and generalization that are usually evaluated in the context of classification, we also need to assess how the algorithm handles two main challenges of the incremental learning: the concept drift and the catastrophic forgetting. However, only catastrophic forgetting is evaluated by the current methodology, where the classifier is evaluated in two scenarios for class addition and expansion. We generalize the methodology by proposing two new scenarios of incremental learning for class inclusion and separation that evaluate the handling of the concept drift. We demonstrate the proposed methodology on the evaluation of three different incremental classifiers, where we show that the proposed methodology provides a more complete and finer evaluation.

Activity-based modelling is a modern agent-based approach to travel demand modelling, in which the transport demand is derived from the agent’s needs to perform certain activities at specific places and times. The agent’s mobility is considered in a broader context, which allows the activity-based models to produce more realistic trip chains, compared to traditional trip-based models. The core of any activity-based model is an activity scheduler – a software component producing sequences of agent’s daily activities interconnected by trips, called activity schedules. Traditionally, activity schedulers used to rely heavily on hard-coded knowledge of transport behaviour experts. We introduce the concept of a Data-Driven Activity Scheduler (DDAS), which replaces numerous expert-designed components and their intricately engineered interactions with a collection of machine learning models. Its architecture is significantly simpler, making it easier to deploy and maintain. This shift towards data-driven, machine learning based approach is possible due to increased availability of mobility-related data. We demonstrate DDAS concept using our own proof-of-concept implementation, perform a rigorous analysis and compare the validity of the resulting model to one of the rule-based alternatives using the Validation Framework for Activity-Based Models (VALFRAM).

Terrain classification is a crucial feature for mobile robots operating across multiple terrains. One way to learn a terrain classifier is to use a stream of labeled proprioceptive data recorded during a terrain traversal. In this paper, we propose a new terrain classifier that combines a feature extraction from a data stream with the long short-term memory (LSTM) network. Features are extracted from the information-sparse data stream by applying a sliding window computing three central moments. The feature sequence is continuously classified by the LSTM network into multiple terrain classes. Furthermore, a modified bagging method is used to deal with a limited and unbalanced training set. In comparison to the previous work on terrain classifiers for a hexapod crawling robot using only servo-drive feedback, the proposed classifier provides continuous classification with the F1 score up to 0.88, and thus provide better results than SVM classifier learned on the same input data.

On-demand transport has been disrupted by Uber and other providers, which are challenging the traditional approach adopted by taxi services. Instead of using fixed passenger pricing and driver payments, there is now the possibility of adaptation to changes in demand and supply. Properly designed, this new approach can lead to desirable tradeoffs between passenger prices, individual driver profits and provider revenue. However, pricing and allocations known as mechanisms are challenging problems falling in the intersection of economics and computer science. In this paper, we develop a general framework to classify mechanisms in ondemand transport. Moreover, we show that data is key to optimizing each mechanism and analyze a dataset provided by a real-world on-demand transport provider. This analysis provides valuable new insights into efficient pricing and allocation in on-demand transport.

Activity-based models, as a specific instance of agent-based models, deal with agents that structure their activity in terms of (daily) activity schedules. An activity schedule consists of a sequence of activity instances, each with its assigned start time, duration and location, together with transport modes used for travel between subsequent activity locations. A critical step in the development of simulation models is validation. Despite the growing importance of activity-based models in modelling transport and mobility, there has been so far no work focusing specifically on statistical validation of such models. In this paper, we propose a six-step Validation Framework for Activity-based Models (VALFRAM) that allows exploiting historical real-world data to assess the validity of activity-based models. The framework compares temporal and spatial properties and the structure of activity schedules against real-world travel diaries and origin-destination matrices. We confirm the usefulness of the framework on three real-world activity-based transport models.

Activity-based models are a specific type of agent-based models widely used in transport and urban planning to generate and study travel demand. They deal with agents that structure their behaviour in terms of daily activity schedules: sequences of activity instances (such as work, sleep or shopping) with assigned start times, durations and locations, and interconnected by trips with assigned transport modes and routes. Despite growing importance of activity-based models in transport modelling, there has been no work focusing specifically on statistical validation of such models so far. In this paper, we propose a six-step Validation Framework for Activity-based Models (VALFRAM) that exploits historical real-world data to quantify the model's validity in terms of a set of numeric metrics. The framework compares the temporal and spatial properties and the structure of modelled activity schedules against real-world origin-destination matrices and travel diaries. We demonstrate the usefulness of the framework on a set of six different activity-based transport models.

In this paper we present CaffeLink an open-source library for Mathematica which is a binding of a well-established Caffe deep learning framework. Caffe is a highly-optimized CUDA accelerated library with focus on convolutional neural networks written in C++ with Python and Matlab bindings. CaffeLink is based upon Mathematica's LibraryLink. It makes accessible most features of Caffe directly from Mathematica environment which includes work with datasets, building networks, training them as well as evaluating them. Here we present an overview of the CaffeLink library with examples on MNIST and ImageNet datasets.

Even though the agent-based simulation modelling has become a standard tool in transport research, current imple- mentations still treat travellers as passive data structures, updated synchronously at infrequent, predefined points in time, thus failing to cover within-the-day decision making and negotiation necessary for cooperative behaviour in a dynamic transport system. Leveraging the fully agent-based modelling approach, we have built large-scale activity-based models of multimodal mobility covering areas up to thousands of square kilometres and simulating populations of up to millions of inhabitants of several European cities. Citizens are represented by autonomous, self-interested agents which schedule and execute their activities (work, shopping, leisure, etc.) and trips in time and space. Indi- vidual decisions are influenced by agent’s demographic attributes and modelled using the data from mobility surveys. The model is statistically validated against origin-destination matrices and travel diary data sets.

In this paper we present a novel algorithm called GPAT (Genetic Programming of Augmenting Topologies) which evolves Genetic Programming (GP) trees in a similar way as a well-established neuro-evolutionary algorithm NEAT (NeuroEvolution of Augmenting Topologies) does. The evolution starts from a minimal form and gradually adds structure as needed. A niching evolutionary algorithm is used to protect individuals of a variable complexity in a single population. Although GPAT is a general approach we employ it mainly to evolve artificial neural networks by means of Hypercube-based indirect encoding which is an approach allowing for evolution of large-scale neural networks having theoretically unlimited size. We perform also experiments for directly encoded problems. The results show that GPAT outperforms both GP and NEAT taking the best of both.

In this paper we propose a new algorithm called HyperGPEFS (HyperGP with Explicit Fitness Sharing). It is based on a HyperNEAT, which is a well-established evolutionary method employing indirect encoding of artificial neural networks. Indirect encoding in HyperNEAT is realized via special function called Compositional and Pattern Producing Network (CPPN), able to describe a neural network of arbitrary size. CPPNs are represented by network structures, which are evolved by means of a slightly modified version of another, well-known algorithm NEAT (NeuroEvolution of Augmenting Topologies). HyperGP is a variant of HyperNEAT, where the CPPNs are optimized by Genetic Programming (GP). Published results reported promising improvement in the speed of convergence. Our approach further extends HyperGP by using fitness sharing to promote a diversity of a population. Here, we thoroughly compare all three algorithms on six different tasks. Fitness sharing demands a definition of a tree distance measure. Among other five, we propose a generalized distance measure which, in conjunction with HyperGPEFS, significantly outperforms HyperNEAT and HyperGP on all, but one testing problems. Although this paper focuses on indirect encoding, the proposed distance measures are generally applicable.

Optimization of neural network topology, weights and neuron transfer functions for given data set and problem is not an easy task. In this article, we focus primarily on building optimal feed-forward neural network classifier for i.i.d. data sets. We apply metalearning principles to the neural network structure and function optimization. We show that diversity promotion, ensembling, self-organization and induction are beneficial for the problem. We combine several different neuron types trained by various optimization algorithms to build a supervised feedforward neural network called Group of Adaptive Models Evolution (GAME). The approach was tested on wide number of benchmark data sets. The experiments show that the combination of different optimization algorithms in the network is the best choice when the performance is averaged over several real-world problems.

In this paper we present neuro-evolution of neural network controllers for mobile agents in a simulated environment. The controller is obtained through evolution of hypercube encoded weights of recurrent neural networks (HyperNEAT). The simulated agent's goal is to find a target in a shortest time interval. The generated neural network processes three different inputs - surface quality, obstacles and distance to the target. A behavior emerged in agents features ability of driving on roads, obstacle avoidance and provides an efficient way of the target search.

In this paper we describe simulation of autonomous robots controlled by recurrent neural networks, which are evolved through indirect encoding using HyperNEAT algorithm. The robots utilize 180 degree wide sensor array. Thanks to the scalability of the neural network generated by HyperNEAT, the sensor array can have various resolution. This would allow to use camera as an input for neural network controller used in real robot. The robots were simulated using software simulation environment. In the experiments the robots were trained to drive with imaximum average speed. Such fitness forces them to learn how to drive on roads and avoid collisions. Evolved neural networks show excellent scalability. Scaling of the sensory input breaks performance of the robots, which should be gained back with re-training of the robot with a different sensory input resolution.

In this paper we present state of the art in the field of meta-optimization. We are looking for possible applications of Inductive Approach in this field. We propose an inductive based optimization strategy combining several optimization algorithms to solve increasingly complex problems.

The visualization of multi-dimensional data is a very important task in computational intelligence methods. This article describes a new algorithm which was originally designed to visualize diversity of population in Evolutionary Algorithms (EAs). Later, we realized its abilities for visualization of any general data sets which makes it suitable for data mining purposes.

This paper focuses on so-called TWEANNs (Toppology and Weight Evolving Artificial Neural Networks). Here, we concentrate on a use of an indirect developmental encoding which is an approach inspired by multi-cellular organisms' development from a single cell. We examine multiple modifications of a known tree-based indirect developmental encoding: the Cellular Encoding. Grammatical Evolution (GE) is employed instead of Genetic Programming (GP) to optimize program trees. GE is advantageous mainly in the way it can handle constraints (as it evolves program trees which conform to a grammar prespecified using a BNF notation). Moreover, we employ GE's inner mechanisms to efficiently encode Neural Network parameters (weights and biases). In this work, we compare three different link selection schemes. The results of our investigations show that our modifications of Cellular Encoding improve the ability to evolve real-valued Artificial Neural Networks.

The visualization is an important technique used in many stages of data mining process. This article deals mostly with visualization for preprocessing purposes. The aim of our approach is to visualize distances (Euclidean or others) between data samples. This can be helpful when taking picture of data clustering. In classification tasks it can be used to select outlayer for removal. In this paper we present a novel way of such visualization which is based on a physical system simulation. It is inspired by intermolecular forces and employs overall energy minimization. This minimization is done via known unconstrained optimization numerical methods such as Steepest Descent, Conjugated Gradients or Quasi-Newton. The proposed algorithm was originally designed and was found useful when interpretting diversity in evolutionary algorithms. Here, we show its properties on well-known datasets Iris and Ecoli.

Evolutionary Algorithms (EAs) are well-known nature-inspired optimization methods. Diversity is an essenial aspect of each EA. It describes the variability of organisms in population. The lack of diversity is common problem - diversity should be preserved in order to evade local extremes (premature convergence). Niching algorithms are modifications of classical EAs. Niching is based on dividing the population into separate subpopulations - it spreads the organisms effectively all over the search space and hence making the overall population diverse. Using niching methods also requires setting of their parameters, which can be very difficult. This paper presents a novel way of diversity visualization based on physical system simulation. This visualization is helpful when designing and tuning niching algorithms but it has also other uses. The visualization will be presented on NEAT - the evolutionary algorithm which optimizes both the topology and the parameters of neural networks.

This article deals with visualization of diversity in Evolutionary Algorithms (EAs) Diversity is an essenial aspect of each EA. It describes the variability of individuals in population. The lack of diversity is common problem - diversity should be preserved in order to evade local extremes (premature convergence). Niching algorithms are modifications of classical EAs. Niching is based on dividing the population into separate subpopulations - it spreads the individuals effectively all over the search space and hence making the overall population diverse. Using niching methods also requires setting of their parameters, which can be very difficult. This paper presents a novel way of diversity visualization based on simulation of intermolecular forces. This visualization is helpful when designing and tuning niching algorithms but it can also be used to visualize any data on which a similarity measure is defined.

EAs can be seen as algorithms which traverse the search-space in a parallel way. Diversity is an essential aspect of each EA. The lack of diversity is a common problem. Diversity should be preserved in order to evade local extremes (premature convergence). Niching EA is based on dividing the population into separate subpopulations - it spreads the organisms effectively all over the search-space and hence making the overall population diverse. Using niching methods also requires setting of their parameters, which can be very difficult. This paper presents a novel way of diversity visualization based on physical system simulation. It is inspired by intermolecular forces and employs overall energy minimization. This minimization is done via known unconstrained optimization numerical methods. The visualization is helpful when designing and tuning niching algorithms, but it has also other uses.

Diversity is one of the most important factors that influences quality of results of computational intelligence methods. Study of diversity is not straightforward, since the diversity usually arises from a multidimensional environment. In this paper we present several methods how to visualize diversity in computational intelligence, namely evolutionary algorithms and artificial neural networks. In evolutionary algorithms, there is a need to visualize diversity of individuals in populations. In neural networks a diversity of neurons is being studied. We designed and used several techniques for visualization of diversities.

This paper deals with creation of models by means of evolutionary algorithms, particularly with maintaining diversity of population using niching methods. Niching algorithms are known for their ability to search for more optima simultaneously. This is done by splitting the population of models into separate species. Species protect promising but yet not fully developed models. Search for more optima at the same time helps to avoid a premature convergence and therefore deals effectively with local optima. Efficiency of two different niching methods is compared on NEAT applied to the neuro-evolution of models.