Brazil

An important and often neglected aspect in probabilistic planning is how to account for different attitudes towards risk in the process. In goal-driven problems, modeled as Shortest Stochastic Path (SSP) problems, risk arises from the uncertainties on future events and how they can lead to goal states. An SSP agent that minimizes the expected accumulated cost is considered a risk-neutral agent, while with a different optimization criterion it could choose between two extreme attitudes: risk-aversion or risk-prone. In this work we consider a Risk Sensitive ssp (called RS-SSP) that uses an expected exponential utility parameterized by the risk factor lambda that is used to define the agent&#39;s risk attitude. Moreover, a lambda-value is feasible if it admits a policy with finite expected cost. There are several algorithms capable of determining an optimal policy for rs-ssps when we fix a feasible value for lambda. However, so far, there has been only one approach to find an extreme lambda feasible i.e., an extreme risk-averse policy. In this work we propose and compare new approaches to finding the extreme feasible lambda value for a given RS-SSP, and to return the corresponding extreme risk-averse policy. Experiments on three benchmark domains show that our proposals outperform previous approach, allowing the solution of larger problems

BRACIS 2020

Finding Feasible Policies for Extreme Risk-Averse Agents in Probabilistic Planning

dual linear programming

risk sensitive mdp

probabilistic planning

The 9th Brazilian Conference on Intelligent Systems (BRACIS), the 16h Encontro Nacional de Inteligência Artificial e Computacional (ENIAC), and the 12th Concurso de Teses e Dissertações em Inteligência Artificial e Computacional (CTDIAC) was held from 20 to 23 October, 2020.

BRACIS originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence - SBIA (22 editions), and the Brazilian Symposium on Neural Networks - SBRN (13 editions). ENIAC is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. CTDIAC is a satellite event whose goals are to disseminate and award the best dissertations and theses (MSc and PhDs concluded in Brazilian universities from June 2018 to May 2020) in Artificial and Computational Intelligence areas.

In this year, the Symposium on Knowledge Discovery, Mining and Learning (KDMiLe) will be held in conjunction with BRACIS 2020. This sysposium aims at integrating researchers, practitioners, developers, students and users to present their research results, to discuss ideas, and to exchange techniques, tools, and practical experiences – related to the Data Mining and Machine Learning areas.

An important and often neglected aspect in probabilistic planning is how to account for different attitudes towards risk in the process. In goal-driven problems, modeled as Shortest Stochastic Path (SSP) problems, risk arises from the uncertainties on future events and how they can lead to goal states. An SSP agent that minimizes the expected accumulated cost is considered a risk-neutral agent, while with a different optimization criterion it could choose between two extreme attitudes: risk-aversion or risk-prone. In this work we consider a Risk Sensitive ssp (called RS-SSP) that uses an expected exponential utility parameterized by the risk factor lambda that is used to define the agent's risk attitude. Moreover, a lambda-value is feasible if it admits a policy with finite expected cost. There are several algorithms capable of determining an optimal policy for rs-ssps when we fix a feasible value for lambda. However, so far, there has been only one approach to find an extreme lambda feasible i.e., an extreme risk-averse policy. In this work we propose and compare new approaches to finding the extreme feasible lambda value for a given RS-SSP, and to return the corresponding extreme risk-averse policy. Experiments on three benchmark domains show that our proposals outperform previous approach, allowing the solution of larger problems

technical paper

Ensembles of classifiers are a very popular type of method for performing classification, due to their usually high predictive accuracy. However, ensembles have two drawbacks. First, ensembles are usually considered a 'black box', non-interpretable type of classification model, mainly because typically there are a very large number of classifiers in the ensemble (and often each classifier in the ensemble is a black-box classifier by itself). This lack of interpretability is an important limitation in application domains where a model's predictions should be carefully interpreted by users, like medicine, law, etc. Second, ensemble methods typically involve many hyper-parameters, and it is difficult for users to select the best settings for those hyper-parameters. In this work we propose an Evolutionary Algorithm (an Estimation of Distribution Algorithm) that addresses both these drawbacks. This algorithm optimizes the hyper-parameter settings of a small ensemble of 5 interpretable classifiers, which allows users to interpret each classifier. 
In our experiments, the ensembles learned by the proposed Evolutionary Algorithm achieved the same level of predictive accuracy as a well-known Random Forest ensemble, but with the benefit of learning interpretable models (unlike Random Forests).

An Evolutionary Algorithm for Learning Interpretable Ensembles of Classifiers

Image classification is a well-established problem in computer vision. Most state-of-the-art models rely on Convolutional Neural Networks to achieve near-human performance in that task. However, CNNs have shown to be susceptible to image manipulation, which undermines the trustability of perception systems. This property is critical, especially in unmanned systems, autonomous vehicles, and scenarios where light cannot be controlled. We investigate the robustness of several Deep-Learning based image recognition models and how the accuracy is affected by several distinct image distortions. The distortions include ill-exposure, low-range image sensors, and common noise types. Furthermore, we also propose and evaluate an image pipeline designed to minimize image distortion before the image classification is performed. Results show that most CNN models are marginally affected by mild miss-exposure and Shot noise. On the one hand, the proposed pipeline can provide significant gain on miss-exposed images. On the other hand, harsh miss-exposure, signal-dependent noise, and impulse noise incur in a high impact on all evaluated models.

A Pipelined Approach to Deal with Image Distortion in Computer Vision

Image Enhancement is one of the most important phases of the image processing system. Contrast Enhancement plays a key role in this step. Histogram Equalization (HE) is one of the main tools used to improve the contrast of an image. However, the use of HE causes an increase in the natural brightness of the image, which is not desirable in many types of applications such as consumer electronics products. This presentation will show a variation of the Differential Evolution metaheuristic algorithm for Contrast Optimization, called DECO, and the results of the evaluation fo the proposed algorithm.

A Differential Evolution Algorithm for Contrast Optimization

Classification is an essential task in the field of MachineLearning, where developing a classifier that minimizes errors on unknown data is one of its central problems. It is known that the analytic centeris a good approximation of the center of mass of the version space that is consistent with the Bayes-optimal decision surface. Therefore, in this work, we propose an evolutionary algorithm, relying on the convexity properties of the version space, that evolves a population of perceptron classifiers in order to find a solution that approximates its analytic cen-ter. Hyperspherical coordinates are used to guarantee feasibility when generating new individuals and enabling exploration to be uniformly distributed through the search space. To evaluate the individuals we consider using a potential function that employs a logarithmic barrier penalty. Experiments were performed on real datasets, and the obtained results indicate concrete possibilities for applying the proposed algorithm for solving practical problems.

An Evolutionary Analytic Center Classifier

In the Multi-Agent Pickup and Delivery (MAPD) problem, agents must process a sequence of tasks that may appear in the system in different time-steps. Commonly, this problem has two parts: (i) task allocation, where the agent receives the appropriate task, and (ii) path planning, where the best path for the agent to perform its task, without colliding with other agents, is defined. In this work, we propose an integer-encoded genetic algorithm for solving the task allocation part of the MAPD problem combined with two-path planning algorithms already known in the literature: the Prioritized Planning and the Improved Conflict-Based Search (ICBS). Computational experiments were carried out with different numbers of agents and the frequency of tasks. The results show that the proposed approach achieves better results for large instances when compared to another technique from the literature.

Solving Multi-Agent Pickup and Delivery Problems Using a Genetic Algorithm

Basically, the presentation consists of contextualizing and motivating the work done, describing the objectives and contributions. Then, the methodology is approached together with some comparisons with the related works. Next, the experiments and results are commented with a brief discussion. By the end, the conclusion and future work are presented. Finally, a short video is presented showing the performance of the best agent of our work.

Improving FIFA Player Agents Decision-Making Architectures based on Convolutional Neural Networks through Evolutionary Techniques

Among the existing techniques to improve the performance of metaheuristics in optimization problems, adaptive parameter control consists in varying one or more parameters of a given metaheuristic according to some indicator of the search conditions. This approach allows metaheuristics to change algorithmic behaviour during the search, and is particularly relevant for the optimization of dynamic problems. In this research we theoretically analyse in which ways the parameters of the ant colony optimization for continuous domains metaheuristic can be adapted, regarding how each parameter influences exploration and exploitation characteristics of the algorithm. Our experimental contributions include validating the colony success rate as a search condition estimator and choosing suitable maps from this estimator to the parameters q and ksi of the algorithm. Beyond that, we compare the performances of three proposed adaptive versions of the base metaheuristic and show the benefits of simultaneously adapting multiple parameters.

On the Multiple Possible Adaptive Mechanisms of the Continuous Ant Colony Optimization

In this paper, we evaluated the Knowledge Base building of Fuzzy Rule-Based Classification Systems (FRBCS) with the purpose of find a balance between the accuracy and interpretability objectives. Regarding to build, we compared two well-known algorithms: Wang-Mendel, to generate the rule base, and NSGA-II, to learn the rules and tuning the membership functions. The Wang-Mendel algorithm was also used to introduce a seed in NSGA-II initial population, in order to increase its quality and improve convergence speed. Taking into account that the automatic building of the fuzzy systems knowledge base is challenger, because of the amount of data available in several real problems, we analysed the impact of data reduction on it. The experiments were carried out with 23 datasets divided into small and medium-large size, and the results showed that the use of genetic learning is suitable to large datasets as well as data reduction, improving the accuracy and interpretability of the FRBCS arising.

Genetic Learning Analysis of Fuzzy Rule-Based Classifications Systems considering Data Reduction

Forecasting photovoltaic (PV) power generation, as in many other time series scenarios, is a challenging task. Most current solutions for time series forecasting are grounded on Machine Learning (ML) algorithms, which usually outperform statistical-based methods. However, solutions based on ML and, more recently, Deep Learning (DL) have been found vulnerable to adversarial attacks throughout their execution. With this in mind, in this work we explore four time series analysis techniques, namely Naive, a baseline technique for time series, Autoregressive Integrated Moving Average (ARIMA), from the statistical field, and Long Short-term Memory (LSTM) and Temporal Convolutional Network (TCN), from the DL family. These techniques are used to forecast the power generation of a PV power plant 15 minutes and 24 hours ahead, having as input only power generation historical data. Two main aspects were analyzed: i) how training size influenced the performance of the forecasting models and ii) how univariate time series data could be modified by an adversarial attack to decrease models' performance through cross-technique transferability. For i), the mentioned methods were used and evaluated with monthly updates. For ii), Fast Gradient Sign Method (FGSM), along with a logistic regression substitute model and past data, were used to perform attacks against DL models at test time. LSTM and TCN decreased the error as the training sample size increased and outperformed Naive and ARIMA models. Adversarial samples were able to reduce the performance of LSTM and TCN, particularly for short-term forecasts.

Photovoltaic Generation Forecast: Model Training and Adversarial Attack Aspects

Word segmentation can contribute to improve the results of natural language processing tasks on several problem domains, including social media sentiment analysis, source code summarization and neural machine translation. Taking the English language as a case study, we fine-tune a Transformer architecture which has been trained through the Pre-trained Distillation (PD) algorithm, while comparing it to previous experiments with recurrent neural networks. We organize datasets and resources from multiple application domains under a unified format, and demonstrate that our proposed architecture has competitive performance and superior cross-domain generalization in comparison with previous approaches for word segmentation in Western languages.

Finding Feasible Policies for Extreme Risk-Averse Agents in Probabilistic Planning

Downloads

Next from BRACIS 2020

An Evolutionary Algorithm for Learning Interpretable Ensembles of Classifiers

Similar lecture

Monte Carlo Tree Search Algorithm for SSPs Under the GUBS Criterion

Finding Feasible Policies for Extreme Risk-Averse Agents in Probabilistic Planning

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from BRACIS 2020

An Evolutionary Algorithm for Learning Interpretable Ensembles of Classifiers

Similar lecture

Monte Carlo Tree Search Algorithm for SSPs Under the GUBS Criterion

Downloads