Search

Reinforcement learning (RL) is a powerful technique that allows agents to learn optimal decision-making policies through interactions with an environment. However, traditional RL algorithms suffer from several limitations such as the need for large amounts of data and long-term credit assignment, i.e. the problem of determining which actions actually produce a certain reward. Recently, Transformers have shown their capacity to address these constraints in this area of learning in an offline setting. This paper proposes a framework that uses Transformers to enhance the training of online off-policy RL agents and address the challenges described above through self-attention. The proposal introduces a hybrid agent with a mixed policy that combines an online off-policy agent with an offline Transformer agent using the Decision Transformer architecture. By sequentially exchanging the experience replay buffer between the agents, the agent's learning training efficiency is improved in the first iterations and so is the training of Transformer-based RL agents in situations with limited data availability or unknown environments. [ABSTRACT FROM AUTHOR]

The aim of this paper is to represent time series in a fuzzy way by means of a piecewise linear segment method that represents the time series in an efficient way. But in addition, this representation collects the uncertainty generated in the process of generation of the segments. This representation can be used, for instance, to compare time series or to query information from these. That is why in this approach we also propose a method to compare different parts of the output, obtaining as result a dissimilarity value representing how close their outputs are. Several tests have been done in order to know if input time series are represented correctly with this method and to know if the comparison process obtains the expected results. The input data selected for the tests is the real data obtained from the electric consumption of three buildings from the campus in the city of Toledo of our university. [ABSTRACT FROM AUTHOR]

This paper presents a new method to automatically obtain linguistic descriptions from a multivariate time series. The first step consists of the induction from the multivariate time series of a fuzzy model called the temporal fuzzy model. Later this fuzzy model is analyzed in order to obtain trends based on the output variable. With this information, the trends for the input variables are also obtained and together with the output variable trends are represented in such a way that the temporal evolution of the trend is stored within the own-trends. In a trend there are always some elements or points of interest that are more relevant than others for generating a relevant final description. These elements are also extracted and stored, making the final description generation process more efficient. An event search process generates the final linguistic description. The events are identified in the trends or in the structure containing the points of interest previously selected. Once an event is identified, new text is added at the end of the linguistic description. The design of the events and the text related to each event is made with the cooperation of experts in the field of application. The presented approach is checked in sports, more concretely, in countermovement jumping. [ABSTRACT FROM AUTHOR]

Intelligent surveillance refers to using Artificial Intelligence techniques in order to improve surveillance and deal with semantic information obtained from low-level security devices. In this context, the use of expert knowledge may offer a more realistic solution when dealing with the design of a surveillance system. In this work, a conceptual framework based on normality analysis to detect abnormal behaviors by means of normality concepts is presented. A normality concept specifies how a certain object should ideally behave in a concrete environment depending on such a concept. The definition of the normal path concept is studied in depth in order to analyze behaviors in an outdoor environment. [ABSTRACT FROM AUTHOR]

Rendering is the process of generating a 2D image from the abstract description of a 3D scene. In spite of the development of new techniques and algorithms, the computational requirements of photorealistic rendering are huge so that it is not possible to render them in real time. In addition, the adequate configuration of rendering quality parameters is very difficult to be done by inexpert users, and they are usually set higher than in fact are needed. This article presents an architecture called MAgArRO to optimize the rendering process in a distributed, noncentralized way through a multiagent solution, by making use of expert knowledge or previous jobs to reduce the final rendering. Experimental results prove that this novel approach offers a promising research line to optimize the rendering of photorealistic images. [ABSTRACT FROM AUTHOR]

The aim of this study was to characterise all the goal scoring patterns during open play (elaborate attacks versus counterattacks) related to zone pitch division and the number of players involved in the 2018 FIFA World Cup in Russia. An Iterative Dichotomiser 3 (ID3) decision tree algorithm was used to classify all the goal scoring patterns (94 goals in 64 matches). The results did not show statistical differences between the type of scoring goal during the 2018 FIFA World Cup (p > 0.05; ES = Moderate). According to the result of the patterns of how the goals were achieved, an ID3 algorithm decision tree with seven classification decision nodes was calculated. Consequently, this study may aid national team coaches for the next World Cup to establish notational analyses and spatial-temporal relations to understand how scoring patterns during open play are related to zone pitch division and the number of players involved. [ABSTRACT FROM AUTHOR]