Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660145
M. Beer, O. Kosheleva, V. Kreinovich
In many practical situations, we know that there is a functional dependence between a quantity $q$ and quantities a1,…, an, but the exact form of this dependence is only known with uncertainty. In some cases, we only know the class of possible functions describing this dependence. In other cases, we also know the probabilities of different functions from this class - i.e., we know the corresponding random field or random process. To solve problems related to such a dependence, it is desirable to be able to simulate the corresponding functions, i.e., to have algorithms that transform simple intervals or simple random variables into functions from the desired class. Many of the real-life dependencies are very complex, requiring a large amount of computation time even if we ignore the uncertainty. So, to make simulation of uncertainty practically feasible, we need to make sure that the corresponding simulation algorithm is as fast as possible. In this paper, we show that for this objective, ideas behind neural networks lead to the known Karhunen-Loevc decomposition and interval field techniques - and also that these ideas help us go - when necessary - beyond these techniques.
{"title":"Uncertainty: Ideas Behind Neural Networks Lead Us Beyond KL- Decomposition and Interval Fields","authors":"M. Beer, O. Kosheleva, V. Kreinovich","doi":"10.1109/SSCI50451.2021.9660145","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660145","url":null,"abstract":"In many practical situations, we know that there is a functional dependence between a quantity $q$ and quantities a1,…, an, but the exact form of this dependence is only known with uncertainty. In some cases, we only know the class of possible functions describing this dependence. In other cases, we also know the probabilities of different functions from this class - i.e., we know the corresponding random field or random process. To solve problems related to such a dependence, it is desirable to be able to simulate the corresponding functions, i.e., to have algorithms that transform simple intervals or simple random variables into functions from the desired class. Many of the real-life dependencies are very complex, requiring a large amount of computation time even if we ignore the uncertainty. So, to make simulation of uncertainty practically feasible, we need to make sure that the corresponding simulation algorithm is as fast as possible. In this paper, we show that for this objective, ideas behind neural networks lead to the known Karhunen-Loevc decomposition and interval field techniques - and also that these ideas help us go - when necessary - beyond these techniques.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126240868","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9659877
Y. Kuroe, Kenya Takeuchi, Y. Maeda
In last decades the reinforcement learning method has attracted a great deal of attention and many studies have been done. However, this method is basically a trial-and-error scheme and it takes much computational time to acquire optimal strategies. Furthermore, optimal strategies may not be obtained for large and complicated problems with many states. To resolve these problems we have proposed the swarm reinforcement learning method, which is developed inspired by the multi-point search optimization methods. The Swarm reinforcement learning method has been extensively studied and its effectiveness has been confirmed for several problems, especially for Markov decision processes where the agents can fully observe the states of environments. In many real-world problems, however, the agents cannot fully observe the environments and they are usually partially observable Markov decision processes (POMDPs). The purpose of this paper is to develop a swarm reinforcement learning method which can deal with POMDPs. We propose a swarm reinforcement learning method based on HQ-learning, which is a hierarchical extension of Q-learning. It is shown through experiments that the proposed method can handle POMDPs and possesses higher performance than that of the original HQ-learning.
{"title":"Swarm Reinforcement Learning Method Based on Hierarchical Q-Learning","authors":"Y. Kuroe, Kenya Takeuchi, Y. Maeda","doi":"10.1109/SSCI50451.2021.9659877","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9659877","url":null,"abstract":"In last decades the reinforcement learning method has attracted a great deal of attention and many studies have been done. However, this method is basically a trial-and-error scheme and it takes much computational time to acquire optimal strategies. Furthermore, optimal strategies may not be obtained for large and complicated problems with many states. To resolve these problems we have proposed the swarm reinforcement learning method, which is developed inspired by the multi-point search optimization methods. The Swarm reinforcement learning method has been extensively studied and its effectiveness has been confirmed for several problems, especially for Markov decision processes where the agents can fully observe the states of environments. In many real-world problems, however, the agents cannot fully observe the environments and they are usually partially observable Markov decision processes (POMDPs). The purpose of this paper is to develop a swarm reinforcement learning method which can deal with POMDPs. We propose a swarm reinforcement learning method based on HQ-learning, which is a hierarchical extension of Q-learning. It is shown through experiments that the proposed method can handle POMDPs and possesses higher performance than that of the original HQ-learning.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"101 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128124911","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9659859
Weili Liu, Yue-jiao Gong, Wei-neng Chen, J. Zhong, Sang-Woon Jean, Jun Zhang
With the proliferation of electric vehicles, the Electric Vehicle Charging Scheduling (EVCS) becomes a critical issue in the modern transportation systems. The EVCS problem in practice usually contains several important but conflicting objectives, such as minimizing the time cost, minimizing the charging expense, and maximizing the final state of charge. To solve the multiobjective EVCS (MOEVCS) problem, the weighted-sum approaches require expertise to predefine the weights, which is inconvenient. Meanwhile, traditional Pareto-based approaches require users to frequently select the result from a large set of trade-off solutions, which is sometimes difficult to make decisions. To address these issues, this paper proposes a Heterogeneous Multiobjective Differential Evolution (HMODE) with four heterogeneous sub-populations. Specially, one is for the multiobjective optimization and the other three are single-objective sub-populations that separately optimize three objectives. These four sub-populations are evolved cooperatively to find better trade-off solutions of MOEVCS. Besides, HMODE introduces an attention mechanism to the knee and bound solutions among non-dominated solutions of the first rank to provide more representative trade-off solutions, which facilitates decision makers to select their preferred results. Experimental results show our proposed HMODE outperforms state-of-the-art methods in terms of selection flexibility and solution quality.
{"title":"Heterogeneous Multiobjective Differential Evolution for Electric Vehicle Charging Scheduling","authors":"Weili Liu, Yue-jiao Gong, Wei-neng Chen, J. Zhong, Sang-Woon Jean, Jun Zhang","doi":"10.1109/SSCI50451.2021.9659859","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9659859","url":null,"abstract":"With the proliferation of electric vehicles, the Electric Vehicle Charging Scheduling (EVCS) becomes a critical issue in the modern transportation systems. The EVCS problem in practice usually contains several important but conflicting objectives, such as minimizing the time cost, minimizing the charging expense, and maximizing the final state of charge. To solve the multiobjective EVCS (MOEVCS) problem, the weighted-sum approaches require expertise to predefine the weights, which is inconvenient. Meanwhile, traditional Pareto-based approaches require users to frequently select the result from a large set of trade-off solutions, which is sometimes difficult to make decisions. To address these issues, this paper proposes a Heterogeneous Multiobjective Differential Evolution (HMODE) with four heterogeneous sub-populations. Specially, one is for the multiobjective optimization and the other three are single-objective sub-populations that separately optimize three objectives. These four sub-populations are evolved cooperatively to find better trade-off solutions of MOEVCS. Besides, HMODE introduces an attention mechanism to the knee and bound solutions among non-dominated solutions of the first rank to provide more representative trade-off solutions, which facilitates decision makers to select their preferred results. Experimental results show our proposed HMODE outperforms state-of-the-art methods in terms of selection flexibility and solution quality.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"56 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115882088","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9659889
Noor Asma Husain, M. Rahim, Huma Chaudhry
The dust, mist, haze, and smokiness of the atmosphere typically degrade images from the light and absorption. These effects have poor visibility, dimmed luminosity, low contrast, and distortion of colour. As a result, restoring a degraded image is difficult, especially in hazy conditions. The image dehazing method focuses on improving the visibility of image details while preserving image colours without causing data loss. Many image dehazing methods achieve the goal of removing haze while also addressing other issues such as oversaturation, colour distortion, and halo artefacts. However, the limitation of haze level rendered these approaches ineffective. A volume of various haze level data is required to demonstrate the efficiency of the image dehazing method in removing haze at all haze levels and obtaining the image's quality. This paper introduced a dynamic scattering coefficient to the dehazing algorithm for determining an applicable visibility range for different haze conditions. These proposed methods improve on the current state-of-the-art dehazing method in terms of image quality measurement results.
{"title":"Different Haze Image Conditions for Single Image Dehazing Method","authors":"Noor Asma Husain, M. Rahim, Huma Chaudhry","doi":"10.1109/SSCI50451.2021.9659889","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9659889","url":null,"abstract":"The dust, mist, haze, and smokiness of the atmosphere typically degrade images from the light and absorption. These effects have poor visibility, dimmed luminosity, low contrast, and distortion of colour. As a result, restoring a degraded image is difficult, especially in hazy conditions. The image dehazing method focuses on improving the visibility of image details while preserving image colours without causing data loss. Many image dehazing methods achieve the goal of removing haze while also addressing other issues such as oversaturation, colour distortion, and halo artefacts. However, the limitation of haze level rendered these approaches ineffective. A volume of various haze level data is required to demonstrate the efficiency of the image dehazing method in removing haze at all haze levels and obtaining the image's quality. This paper introduced a dynamic scattering coefficient to the dehazing algorithm for determining an applicable visibility range for different haze conditions. These proposed methods improve on the current state-of-the-art dehazing method in terms of image quality measurement results.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"52 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115887687","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660102
G. Castellano, Andrea Esposito, Marco Mirizio, Graziano Montanaro, G. Vessio
Dementia is one of the most common diseases in the elderly and a leading cause of mortality and disability. In recent years, a research effort has been made to develop computer aided diagnosis tools based on machine (deep) learning models fed with neuroimaging data. However, while much work has been done on MRI imaging, very little attention has been paid on amyloid PETs, which have been recently recognized to be a promising and powerful biomarker of neurodegeneration. In this paper, we contribute to this less explored research area by proposing a 3D Convolutional Neural Network aimed at detecting dementia based on amyloid PET scans. An experiment performed on the recently released OASIS-3 dataset, which provides the community with a new benchmark to advance this line of research further, yielded very promising results and provided new evidence on the effectiveness of amyloid PET.
{"title":"Detection of Dementia Through 3D Convolutional Neural Networks Based on Amyloid PET","authors":"G. Castellano, Andrea Esposito, Marco Mirizio, Graziano Montanaro, G. Vessio","doi":"10.1109/SSCI50451.2021.9660102","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660102","url":null,"abstract":"Dementia is one of the most common diseases in the elderly and a leading cause of mortality and disability. In recent years, a research effort has been made to develop computer aided diagnosis tools based on machine (deep) learning models fed with neuroimaging data. However, while much work has been done on MRI imaging, very little attention has been paid on amyloid PETs, which have been recently recognized to be a promising and powerful biomarker of neurodegeneration. In this paper, we contribute to this less explored research area by proposing a 3D Convolutional Neural Network aimed at detecting dementia based on amyloid PET scans. An experiment performed on the recently released OASIS-3 dataset, which provides the community with a new benchmark to advance this line of research further, yielded very promising results and provided new evidence on the effectiveness of amyloid PET.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115943552","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660108
Vahideh Reshadat, A. Akçay, Kalliopi Zervanou, Yingqian Zhang, Eelco de Jong
Non-transparent shipping processes of transporting goods with special handling needs (special cargoes) have resulted in inefficiency in the airfreight industry. Special cargo ontology elicits, structures, and stores domain knowledge and represents the domain concepts and relationship between them in a machine-readable format. In this paper, we proposed an ontology population pipeline for the special cargo domain, and as part of the ontology population task, we investigated how to build an efficient information extraction model from low-resource domains based on available domain data for industry use cases. For this purpose, a model is designed for extracting and classifying instances of different relation types between each concept pair. The model is based on a relation representation learning approach built upon a Hierarchical Attention-based Multi-task architecture in the special cargo domain. The results of experiments show that the model could represent the complex semantic information of the domain, and tasks initialized with these representations achieve promising results.
{"title":"Relation Representation Learning for Special Cargo Ontology","authors":"Vahideh Reshadat, A. Akçay, Kalliopi Zervanou, Yingqian Zhang, Eelco de Jong","doi":"10.1109/SSCI50451.2021.9660108","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660108","url":null,"abstract":"Non-transparent shipping processes of transporting goods with special handling needs (special cargoes) have resulted in inefficiency in the airfreight industry. Special cargo ontology elicits, structures, and stores domain knowledge and represents the domain concepts and relationship between them in a machine-readable format. In this paper, we proposed an ontology population pipeline for the special cargo domain, and as part of the ontology population task, we investigated how to build an efficient information extraction model from low-resource domains based on available domain data for industry use cases. For this purpose, a model is designed for extracting and classifying instances of different relation types between each concept pair. The model is based on a relation representation learning approach built upon a Hierarchical Attention-based Multi-task architecture in the special cargo domain. The results of experiments show that the model could represent the complex semantic information of the domain, and tasks initialized with these representations achieve promising results.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"232 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115964855","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660177
Sophie Noiret, J. Lumetzberger, M. Kampel
Discriminatory practices involving AI -driven police work have been the subject of much controversies in the past few years, with algorithms such as COMPAS, PredPol and ShotSpotter being accused of unfairly impacting minority groups. At the same time, the issues of fairness in machine learning, and in particular in computer vision, have been the subject of a growing number of academic works. In this paper, we examine how these area intersect. We provide information on how these practices have come to exist and the difficulties in alleviating them. We then examine three applications currently in development to understand what risks they pose to fairness and how those risks can be mitigated.
{"title":"Bias and Fairness in Computer Vision Applications of the Criminal Justice System","authors":"Sophie Noiret, J. Lumetzberger, M. Kampel","doi":"10.1109/SSCI50451.2021.9660177","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660177","url":null,"abstract":"Discriminatory practices involving AI -driven police work have been the subject of much controversies in the past few years, with algorithms such as COMPAS, PredPol and ShotSpotter being accused of unfairly impacting minority groups. At the same time, the issues of fairness in machine learning, and in particular in computer vision, have been the subject of a growing number of academic works. In this paper, we examine how these area intersect. We provide information on how these practices have come to exist and the difficulties in alleviating them. We then examine three applications currently in development to understand what risks they pose to fairness and how those risks can be mitigated.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"81 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134153347","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660041
S. Holly, Astrid Nieße
Fitness landscape analysis is often employed to quantify the properties of optimization problems and hence gain a better understanding of these problems and the behavior of the algorithms applied to them. The calculation of various landscape features requires complete knowledge of the boundaries and constraints of the entire search space. Many real-world applications of distributed optimization exhibit an inherent domain decomposition, i.e., the decision variables for a cooperative search are in the hands of multiple actors. Thus, knowledge about the overall search space - likewise distributed - is not available at a central location. In this paper, we propose an approach for distributed computation and subsequent composition of fitness landscape features. We evaluate the approach with a set of well-known continuous benchmark functions and examine the features for correlation with algorithm performance and their suitability for feature-based algorithm parameterization. The results show that the distributedly computed features provide useful insights into the nature of the problems and that especially the heterogeneity of the sub-search spaces is a relevant factor in the optimized design of the exchange mechanisms of distributed heuristics.
{"title":"Distributed fitness landscape analysis for cooperative search with domain decomposition","authors":"S. Holly, Astrid Nieße","doi":"10.1109/SSCI50451.2021.9660041","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660041","url":null,"abstract":"Fitness landscape analysis is often employed to quantify the properties of optimization problems and hence gain a better understanding of these problems and the behavior of the algorithms applied to them. The calculation of various landscape features requires complete knowledge of the boundaries and constraints of the entire search space. Many real-world applications of distributed optimization exhibit an inherent domain decomposition, i.e., the decision variables for a cooperative search are in the hands of multiple actors. Thus, knowledge about the overall search space - likewise distributed - is not available at a central location. In this paper, we propose an approach for distributed computation and subsequent composition of fitness landscape features. We evaluate the approach with a set of well-known continuous benchmark functions and examine the features for correlation with algorithm performance and their suitability for feature-based algorithm parameterization. The results show that the distributedly computed features provide useful insights into the nature of the problems and that especially the heterogeneity of the sub-search spaces is a relevant factor in the optimized design of the exchange mechanisms of distributed heuristics.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"39 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134239107","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9659552
Abhiroop Ghosh, Yashesh D. Dhebar, Ritam Guha, K. Deb, S. Nageshrao, Ling Zhu, E. Tseng, Dimitar Filev
The recent years have witnessed a surge in application of deep neural networks (DNNs) and reinforcement learning (RL) methods to various autonomous control systems and game playing problems. While they are capable of learning from real-world data and produce adequate actions to various state conditions, their internal complexity does not allow an easy way to provide an explanation for their actions. In this paper, we generate state-action pair data from a trained DNN/RL system and employ a previously proposed nonlinear decision tree (NLDT) framework to decipher hidden simplistic rule sets that interpret the working of DNN/RL systems. The complexity of the rule sets are controllable by the user. In essence, the inherent bilevel optimization procedure that finds the NLDTs is capable of reducing the complexities of the state-action logic to a minimalist and intrepretable level. Demonstrating the working principle of the NLDT method to a revised mountain car control problem, this paper applies the methodology to the lane changing problem involving six critical cars in front and rear in left, middle, and right lanes of a pilot car. NLDTs are derived to have simplistic relationships of 12 decision variables involving relative distances and velocities of the six critical cars. The derived analytical decision rules are then simplified further by using a symbolic analysis tool to provide English-like interpretation of the lane change problem. This study makes a scratch to the issue of interpretability of modern machine learning based tools and it now deserves further attention and applications to make the overall approach more integrated and effective.
{"title":"Interpretable AI Agent Through Nonlinear Decision Trees for Lane Change Problem","authors":"Abhiroop Ghosh, Yashesh D. Dhebar, Ritam Guha, K. Deb, S. Nageshrao, Ling Zhu, E. Tseng, Dimitar Filev","doi":"10.1109/SSCI50451.2021.9659552","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9659552","url":null,"abstract":"The recent years have witnessed a surge in application of deep neural networks (DNNs) and reinforcement learning (RL) methods to various autonomous control systems and game playing problems. While they are capable of learning from real-world data and produce adequate actions to various state conditions, their internal complexity does not allow an easy way to provide an explanation for their actions. In this paper, we generate state-action pair data from a trained DNN/RL system and employ a previously proposed nonlinear decision tree (NLDT) framework to decipher hidden simplistic rule sets that interpret the working of DNN/RL systems. The complexity of the rule sets are controllable by the user. In essence, the inherent bilevel optimization procedure that finds the NLDTs is capable of reducing the complexities of the state-action logic to a minimalist and intrepretable level. Demonstrating the working principle of the NLDT method to a revised mountain car control problem, this paper applies the methodology to the lane changing problem involving six critical cars in front and rear in left, middle, and right lanes of a pilot car. NLDTs are derived to have simplistic relationships of 12 decision variables involving relative distances and velocities of the six critical cars. The derived analytical decision rules are then simplified further by using a symbolic analysis tool to provide English-like interpretation of the lane change problem. This study makes a scratch to the issue of interpretability of modern machine learning based tools and it now deserves further attention and applications to make the overall approach more integrated and effective.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133919009","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-12-05DOI: 10.1109/SSCI50451.2021.9660095
Jenna N. Iorio, R. Regis
The Constrained Accelerated Random Search (CARS) algorithm is a stochastic search method that converges with probability 1 to the global minimum of a constrained black-box optimization problem under certain conditions. CARS randomly selects its sample point from a box centered at the current best solution and adjusts the size of this box depending on whether the point yields an improvement in constraint violation or feasible objective function value. For computationally expensive problems, the CARS- RBF algorithm that uses Radial Basis Function (RBF) surrogates was proposed. Numerical experiments showed the effectiveness of CARS and CARS-RBF compared to alternatives on many test problems. However, both algorithms require a feasible starting point. This paper extends CARS and CARS-RBF to handle constrained black-box optimization problems when a feasible starting point is not available. The extended algorithms begin by minimizing a measure of constraint violation to find a feasible solution and then they search for the global minimum until the computational budget is reached. The algorithms were tested on 19 benchmark problems and on a 12-D engineering optimization problem with 68 black-box constraints where none of the initial points are guaranteed to be feasible. CARS outperformed Constrained Pure Random Search (CPRS) and the ISRES and jDE evolutionary algorithms on the test problems, and CARS-RBF is generally an improvement over CARS. Furthermore, CARS-RBF outperformed other methods including RBF -assisted CPRS and the COBYLA trust region method and it compared favorably with constrained EGO.
约束加速随机搜索(CARS)算法是一种随机搜索方法，在一定条件下以概率1收敛于约束黑箱优化问题的全局最小值。CARS从以当前最佳解为中心的方框中随机选择样本点，并根据该点是否在约束违反或可行目标函数值方面有所改善来调整该方框的大小。针对计算量大的问题，提出了基于径向基函数(RBF)的CARS- RBF算法。数值实验证明了CARS和CARS- rbf算法在许多测试问题上的有效性。然而，这两种算法都需要一个可行的起始点。本文扩展了CARS和CARS- rbf来处理无可行起点时的约束黑盒优化问题。扩展算法首先最小化约束违反量以找到可行解，然后搜索全局最小值，直到达到计算预算。这些算法在19个基准问题和一个具有68个黑盒约束的12维工程优化问题上进行了测试，其中没有一个初始点保证是可行的。CARS在测试问题上优于约束纯随机搜索(Constrained Pure Random Search, CPRS)、ISRES和jDE进化算法，CARS- rbf总体上是对CARS的改进。此外，CARS-RBF优于RBF辅助CPRS和COBYLA信任域方法，且优于约束EGO。
{"title":"Accelerated Random Search for Black-Box Constraint Satisfaction and Optimization","authors":"Jenna N. Iorio, R. Regis","doi":"10.1109/SSCI50451.2021.9660095","DOIUrl":"https://doi.org/10.1109/SSCI50451.2021.9660095","url":null,"abstract":"The Constrained Accelerated Random Search (CARS) algorithm is a stochastic search method that converges with probability 1 to the global minimum of a constrained black-box optimization problem under certain conditions. CARS randomly selects its sample point from a box centered at the current best solution and adjusts the size of this box depending on whether the point yields an improvement in constraint violation or feasible objective function value. For computationally expensive problems, the CARS- RBF algorithm that uses Radial Basis Function (RBF) surrogates was proposed. Numerical experiments showed the effectiveness of CARS and CARS-RBF compared to alternatives on many test problems. However, both algorithms require a feasible starting point. This paper extends CARS and CARS-RBF to handle constrained black-box optimization problems when a feasible starting point is not available. The extended algorithms begin by minimizing a measure of constraint violation to find a feasible solution and then they search for the global minimum until the computational budget is reached. The algorithms were tested on 19 benchmark problems and on a 12-D engineering optimization problem with 68 black-box constraints where none of the initial points are guaranteed to be feasible. CARS outperformed Constrained Pure Random Search (CPRS) and the ISRES and jDE evolutionary algorithms on the test problems, and CARS-RBF is generally an improvement over CARS. Furthermore, CARS-RBF outperformed other methods including RBF -assisted CPRS and the COBYLA trust region method and it compared favorably with constrained EGO.","PeriodicalId":255763,"journal":{"name":"2021 IEEE Symposium Series on Computational Intelligence (SSCI)","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133979161","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：