Search

We focus on the important problem of emergency evacuation, which clearly could benefit from reinforcement learning that has been largely unaddressed. Emergency evacuation is a complex task which is difficult to solve with reinforcement learning, since an emergency situation is highly dynamic, with a lot of changing variables and complex constraints that makes it difficult to train on. In this paper, we propose the first fire evacuation environment to train reinforcement learning agents for evacuation planning. The environment is modelled as a graph capturing the building structure. It consists of realistic features like fire spread, uncertainty and bottlenecks. We have implemented the environment in the OpenAI gym format, to facilitate future research. We also propose a new reinforcement learning approach that entails pretraining the network weights of a DQN based agents to incorporate information on the shortest path to the exit. We achieved this by using tabular Q-learning to learn the shortest path on the building model's graph. This information is transferred to the network by deliberately overfitting it on the Q-matrix. Then, the pretrained DQN model is trained on the fire evacuation environment to generate the optimal evacuation path under time varying conditions. We perform comparisons of the proposed approach with state-of-the-art reinforcement learning algorithms like PPO, VPG, SARSA, A2C and ACKTR. The results show that our method is able to outperform state-of-the-art models by a huge margin including the original DQN based models. Finally, we test our model on a large and complex real building consisting of 91 rooms, with the possibility to move to any other room, hence giving 8281 actions. We use an attention based mechanism to deal with large action spaces. Our model achieves near optimal performance on the real world emergency environment., 21 pages, 14 figures, 4 tables

MedAI: Transparency in Medical Image Segmentation is a challenge held for the first time at the Nordic AI Meet that focuses on medical image segmentation and transparency in machine learning (ML)-based systems. We propose three tasks to meet specific gastrointestinal image segmentation challenges collected from experts within the field, including two separate segmentation scenarios and one scenario on transparent ML systems. The latter emphasizes the need for explainable and interpretable ML algorithms. We provide a development dataset for the participants to train their ML models, tested on a concealed test dataset.

Gathering information of aquatic life is often based on timeconsumingmethods utilizing video feeds. It would be beneficialto capture more information cost-effectively from video feeds.Video based object detection has an ability to achieve this.Recent research has shown promising results with the use ofYOLO for object detection of fish. As underwater conditionscan be difficult and thus fish species are hard to discriminate.This study proposes a hierarchical structure-based YOLO Fishalgorithm in both the classification and the dataset to gainvaluable information. With the use of hierarchical classificationand other techniques. YOLO Fish is a state-of-the-art objectdetector on Nordic fish species, with an mAP of 91.8%. Thealgorithm has an inference time of 26.4 ms, fast enough torun on real-time video on the high-end GPU Tesla V100.

Authors accepted manuscript Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Q-learning is one of the most well-known Reinforcement Learning algorithms. There have been tremendous efforts to develop this algorithm using neural networks. Bootstrapped Deep Q-Learning Network is amongst them. It utilizes multiple neural network heads to introduce diversity into Q-learning. Diversity can sometimes be viewed as the amount of reasonable moves an agent can take at a given state, analogous to the definition of the exploration ratio in RL. Thus, the performance of Bootstrapped Deep Q-Learning Network is deeply connected with the level of diversity within the algorithm. In the original research, it was pointed out that a random prior could improve the performance of the model. In this article, we further explore the possibility of replacing priors with noise and sample the noise from a Gaussian distribution to introduce more diversity into this algorithm. We conduct our experiment on the Atari benchmark and compare our algorithm to both the original and other related algorithms. The results show that our modification of the Bootstrapped Deep Q-Learning algorithm achieves significantly higher evaluation scores across different types of Atari games. Thus, we conclude that replacing priors with noise can improve Bootstrapped Deep Q-Learning’s performance by ensuring the integrity of diversities.

MapAI: Precision in Building Segmentation is a competition arranged with the Norwegian Artificial Intelligence Research Consortium (NORA) in collaboration with Centre for Artificial Intelligence Research at the University of Agder (CAIR), the Norwegian Mapping Authority, AI:Hub, Norkart, and the Danish Agency for Data Supply and Infrastructure. The competition will be held in the fall of 2022. It will be concluded at the Northern Lights Deep Learning conference focusing on the segmentation of buildings using aerial images and laser data. We propose two different tasks to segment buildings, where the first task can only utilize aerial images, while the second must use laser data (LiDAR) with or without aerial images. Furthermore, we use IoU and Boundary IoU to properly evaluate the precision of the models, with the latter being an IoU measure that evaluates the results' boundaries. We provide the participants with a training dataset and keep a test dataset for evaluation.

Author's accepted manuscript © 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Detection of anomalies based on smart meter data is crucial to identify potential risks and unusual events at an early stage. The available advanced information and communicating platform and computational capability renders smart grid prone to attacks with extreme social, financial and physical effects. The smart network enables energy management of smart appliances contributing support for ancillary services. Cyber threats could affect operation of smart appliances and hence the ancillary services, which might lead to stability and security issues. In this work, an overview is presented of different methods used in anomaly detection, performance evaluation of 3 models, the k-Nearest Neighbor, local outlier factor and isolated forest on recorded smart meter data from urban area and rural region

Electrical demand forecasting is essential for power generation capacity planning and integrating environment-friendly energy sources. In addition, load predictions will help in developing demand-side management in coordination with renewable power generation. Meteorological conditions influence urban area load pattern; therefore, it is vital to include weather parameters for load predictions. Machine Learning algorithms can effectively be used for electrical load predictions considering impact of external parameters. This paper explores and compares the basic Recurrent Neural Networks (RNN); Simple Recurrent Neural Networks (Vanilla RNN), Gated Recurrent Units (GRU), and Long Short-Term Memory networks (LSTM). Vanilla RNNs are fully connected neural networks where the output from the previous time step is being fed to the next time step. GRUs are networks with a gating mechanism: a forget gate. LSTM networks also, in addition to a forget gate, include an output gate. Even though the recurrent structure in itself is robust for efficient forecasting, pre-processing of data (including load, weather) is important to enhance the performance. Principal Component Analysis (PCA) reduces and extracts the main components of available data. This work shows that PCA improves the performance of RNNs with use of weather parameters. The historical electrical load dataset from Sydney region is used to test the load forecasting using these techniques considering meteorological parameters. Through load forecasting, it is observed that for the 30 minutes predictions, GRU trained with a reduced number of principal components performs best for a typical period with a mean absolute percentage error (MAPE) of 0.74%.

Load forecasting is the most fundamental application in Smart-Grid, which provides essential input to Demand Response, Topology Optimization and Abnormally Detection, facilitating the integration of intermittent clean energy sources. In this work, several regression tools are analyzed using larger datasets for urban area electrical load forecasting. The regression tools which are used are Random Forest Regressor, k-Nearest Neighbour Regressor and Linear Regressor. This work explores the use of regression tool for regional electric load forecasting by correlating lower distinctive categorical level (season, day of the week) and weather parameters. The regression analysis has been done on continuous time basis as well as vertical time axis approach. The vertical time approach is considering a sample time period (e.g seasonally and weekly) of data for four years and has been tested for the same time period for the consecutive year. This work has uniqueness in electrical demand forecasting using regression tools through vertical approach and it also considers the impact of meteorological parameters. This vertical approach uses less amount of data compare to continuous time-series as well as neural network techniques. A correlation study, where both the Pearson method and visual inspection, of the vertical approach depicts meaningful relation between pre-processing of data, test methods and results, for the regressors examined through Mean Absolute Percentage Error (MAPE). By examining the structure of various regressors they are compared for the lowest MAPE. Random Forest Regressor provides better short-term load prediction (30 min) and kNN offers relatively better long-term load prediction (24 h).

Physical inactivity increases the risk of many adverse health conditions, including the world’s major non-communicable diseases, such as coronary heart disease, type 2 diabetes, and breast and colon cancers, shortening life expectancy. There are minimal medical care and personal trainers’ methods to monitor a patient’s actual physical activity types. To improve activity monitoring, we propose an artificial-intelligence-based approach to classify the physical movement activity patterns. In more detail, we employ two deep learning (DL) methods, namely a deep feed-forward neural network (DNN) and a deep recurrent neural network (RNN) for this purpose. We evaluate the proposed models on two physical movement datasets collected from several volunteers who carried tri-axial accelerometer sensors. The first dataset is from the UCI machine learning repository, which contains 14 different activities-of-daily-life (ADL) and is collected from 16 volunteers who carried a single wrist-worn tri-axial accelerometer. The second dataset includes ten other ADLs and is gathered from 8 volunteers who placed the sensors on their hips. Our experiment results show that the RNN model provides the accuracy performance compared to the state-of-the-art methods in classifying the fundamental movement patterns with an overall accuracy of 84.89% and an overall F1-score of 82.56%. Our results indicate that the proposed method will provide the medical doctors and trainers a promising way to precisely track and understand a patient’s physical activities for better treatment.

This paper introduces a novel approach to representing continuous inputs in Tsetlin Machines (TMs). Instead of using one Tsetlin Automaton (TA) for every unique threshold found when Booleanizing continuous input, we employ two Stochastic Searching on the Line (SSL) automata to learn discriminative lower and upper bounds. The two resulting Boolean features are adapted to the rest of the clause by equipping each clause with its own team of SSLs, which update the bounds during the learning process. Two standard TAs finally decide whether to include the resulting features as part of the clause. In this way, only four automata altogether represent one continuous feature (instead of potentially hundreds of them). We evaluate the performance of the new scheme empirically using five datasets, along with a study of interpretability. On average, TMs with SSL feature representation use 4.3 times fewer literals than the TM with static threshold-based features. Furthermore, in terms of average memory usage and F1-Score, our approach outperforms simple Multi-Layered Artificial Neural Networks, Decision Trees, Support Vector Machines, K-Nearest Neighbor, Random Forest, Gradient Boosted Trees (XGBoost), and Explainable Boosting Machines (EBMs), as well as the standard and real-value weighted TMs. Our approach further outperforms Neural Additive Models on Fraud Detection and StructureBoost on CA-58 in terms of the Area Under Curve while performing competitively on COMPAS.

Aspect-based sentiment analysis (ABSA) aims at identifying fine-grained polarity of opinion associated with a given aspect word. Several existing articles demonstrated promising ABSA accuracy using positional embedding to show the relationship between an aspect word and its context. In most cases, the positional embedding depends on the distance between the aspect word and the remaining words in the context, known as the position index sequence. However, these techniques usually employ both complex preprocessing approaches with additional trainable positional embedding and complex architectures to obtain the state-of-the-art performance. In this paper, we simplify preprocessing by including polarity lexicon replacement and masking techniques that carry the information of the aspect word’s position and eliminate the positional embedding. We then adopt a novel and concise architecture using two Bidirectional GRU along with an attention layer to classify the aspect based on its context words. Experiment results show that the simplified preprocessing and the concise architecture significantly improve the accuracy of the publicly available ABSA datasets, obtaining 81.37%, 75.39%, 80.88%, and 89.30% in restaurant 14, laptop 14, restaurant 15, and restaurant 16 respectively.

Tsetlin Machines (TM) use finite state machines for learning and propositional logic to represent patterns. The resulting pattern recognition approach captures information in the form of conjunctive clauses, thus facilitating human interpretation. In this work, we propose a TM-based approach to three common natural language processing (NLP) tasks, namely, sentiment analysis, semantic relation categorization and identifying entities in multi-turn dialogues. By performing frequent itemset mining on the TM-produced patterns, we show that we can obtain a global and a local interpretation of the learning, one that mimics existing rule-sets or lexicons. Further, we also establish that our TM based approach does not compromise on accuracy in the quest for interpretability, via comparison with some widely used machine learning techniques. Finally, we introduce the idea of a relational TM, which uses a logic-based framework to further extend the interpretability.

The Tsetlin Machine (TM) is a recent interpretable machine learning algorithm that requires relatively modest computational power, yet attains competitive accuracy in several benchmarks. TMs are inherently binary; however, many machine learning problems are continuous. While binarization of continuous data through brute-force thresholding has yielded promising accuracy, such an approach is computationally expensive and hinders extrapolation. In this paper, we address these limitations by standardizing features to support scale shifts in the transition from training data to real-world operation, typical for e.g. forecasting. For scalability, we employ sampling to reduce the number of binarization thresholds, relying on stratification to minimize loss of accuracy. We evaluate the approach empirically using two artificial datasets before we apply the resulting TM to forecast dengue outbreaks in the Philippines using the spatiotemporal properties of the data. Our results show that the loss of accuracy due to threshold sampling is insignificant. Furthermore, the dengue outbreak forecasts made by the TM are more accurate than those obtained by Support Vector Machines (SVMs), Decision Trees (DTs), and several multi-layered Artificial Neural Networks (ANNs), both in terms of forecasting precision and Fl-score.

Tsetlin machines (TMs) are a promising approach to machine learning that uses Tsetlin Automata to produce patterns in propositional logic, leading to binary (hard) classifications. In many applications, however, one needs to know the confidence of classifications, e.g. to facilitate risk management. In this paper, we propose a novel scheme for measuring TM confidence based on the logistic function, calculated from the propositional logic patterns that match the input. We then use this scheme to trade off precision against recall, producing area under receiver operating characteristic curves (AUC) for TMs. Empirically, using four real-world datasets, we show that AUC is a more sensitive measure of TM performance compared to Accuracy. Further, the AUC-based evaluations show that the TM performs on par or better than widely used machine learning algorithms. We thus believe our scheme will make the TM more suitable for use in decision support, where the user needs to inspect and validate predictions, in particular, those being uncertain.

In this paper, we propose a generic emergency detection system using only the sound produced in the environment. For this task, we employ multiple audio feature extraction techniques like the mel-frequency cepstral coefficients, gammatone frequency cepstral coefficients, constant Q-transform and chromagram. After feature extraction, a deep convolutional neural network (CNN) is used to classify an audio signal as a potential emergency situation or not. The entire model is based on our previous work that sets the new state of the art in the environment sound classification (ESC) task (Our paper is under review in the IEEE/ACM Transactions on Audio, Speech and Language Processing and also available here https://arxiv.org/abs/1908.11219.) We combine the benchmark ESC datasets: UrbanSound8K and ESC-50 (ESC-10 is a subset of ESC-50) and reduce the problem to a binary classification problem. This is done by aggregating sound classes such as sirens, fire crackling, glass breaking, gunshot as the emergency class and others as normal. Even though there are only two classes to distinguish, they are highly imbalanced. To overcome this difficulty, we introduce class weights in calculating the loss while training the model. Our model is able to achieve \(99.56\%\) emergency detection accuracy.

Deep reinforcement learning is considered promising for many energy cost optimization tasks in smart buildings. How-ever, agent learning, in this context, is sometimes unstable and unpredictable, especially when the environments are complex. In this paper, we examine deep Reinforcement Learning (RL) algorithms developed for game play applied to a battery control task with an energy cost optimization objective. We explore how agent behavior and hyperparameters can be analyzed in a simplified environment with the goal of modifying the algorithms for increased stability. Our modified Deep Deterministic Policy Gradient (DDPG) agent is able to perform consistently close to the optimum over multiple training sessions with a maximum cost reduction of 25 % and an average cost reduction of 99 % of the maximum in the simplified environment. DDPG is an actor-critic RL algorithm consisting of four neural networks - the actor and critic, main and target, networks. When environment complexity is increased, the DDPG agent performance decreases and a modified Twin Delayed DDPG (TD3) agent is utilized to achieve an average of 99.9 % of the optimal result. The TD3 algorithm uses two main critic networks to avoid known value overestimation bias.

Despite significant effort, building models that are both interpretable and accurate is an unresolved challenge for many pattern recognition problems. In general, rule-based and linear models lack accuracy, while deep learning interpretability is based on rough approximations of the underlying inference. Using a linear combination of conjunctive clauses in propositional logic, Tsetlin Machines (TMs) have shown competitive performance on diverse benchmarks. However, to do so, many clauses are needed, which impacts interpretability. Here, we address the accuracy-interpretability challenge in machine learning by equipping the TM clauses with integer weights. The resulting Integer Weighted TM (IWTM) deals with the problem of learning which clauses are inaccurate and thus must team up to obtain high accuracy as a team (low weight clauses), and which clauses are sufficiently accurate to operate more independently (high weight clauses). Since each TM clause is formed adaptively by a team of Tsetlin Automata, identifying effective weights becomes a challenging online learning problem. We address this problem by extending each team of Tsetlin Automata with a stochastic searching on the line (SSL) automaton. In our novel scheme, the SSL automaton learns the weight of its clause in interaction with the corresponding Tsetlin Automata team, which, in turn, adapts the composition of the clause by the adjusting weight. We evaluate IWTM empirically using five datasets, including a study of interpetability. On average, IWTM uses 6.5 times fewer literals than the vanilla TM and 120 times fewer literals than a TM with real-valued weights. Furthermore, in terms of average F1-Score, IWTM outperforms simple Multi-Layered Artificial Neural Networks, Decision Trees, Support Vector Machines, K-Nearest Neighbor, Random Forest, XGBoost, Explainable Boosting Machines, and standard and real-value weighted TMs., 20 pages, 10 figures