16 August 2022 in IEEE Access
In this paper, we propose a novel curiosity-based learning algorithm for Multi-agent Reinforcement Learning (MARL) to attain efficient and effective decision-making. We employ the centralized training with decentralized execution framework (CTDE) and consider that each agent has knowledge of the prior action distribution of others. To quantify the difference in agents’ knowledge, curiosity, we introduce conditional mutual information (CMI) regularization and use the amount of information for updating decision-making policy. Then, to deploy these learning frameworks in a large-scale MARL setting while acquiring high sample efficiency, we consider a Kullback-Leibler (KL) divergence-based prioritization of experiences. We evaluate the effectiveness of the proposed algorithm in three different levels of StarCraft Multi Agent Challenge (SMAC) scenarios using the PyMARL framework. The simulation-based performance analysis shows that the proposed technique significantly improves the test win rate compared to various state-of-the-art MARL benchmarks, such as the Optimistically Weighted Monotonic Value Function Factorization (OW_QMIX) and Learning Individual Intrinsic Reward (LIIR).
22 March 2022 in IEEE Access
Image classification research is one of the fields continuously studied in the computer vision domain, and several related studies have been actively conducted until recently. However, a limit exists regarding the prediction performance of real-world datasets due to the data imbalance problem between classes. Data augmentation through artificial sample generation for minority classes is one of the methods used to overcome this limitation. Among the various oversampling methods, we propose the feature dictionary-based generative model for the oversampling method. Feature dictionaries are built through the pretrained feature extractor, and the proposed generative model synthesizes artificial samples based on the dictionary. Class-to-class balanced training can be conducted by fine-tuning the classifier as additional data for the minority class. We experiment by applying the proposed framework to the fashion dataset, which has an extreme class imbalance. The experimental results demonstrate that the proposed model achieved the highest top-1 performance on various public fashion datasets. In addition, we analyze the number of samples in the dictionary and test the effectiveness of the elements that comprise the proposed model using various ablation studies.
28 April 2022 in IEEE Access
DNDevarani Devi Ningombam
+1
Although recent years witnessed notable success for a cooperative setting in multi-agent reinforcement learning (MARL), efficient explorations are still challenging primarily due to the complex dynamics of inter-agent interactions constituting the high dimension of action spaces. For an efficient exploration, it is necessary to quantify influences that can represent interactions among agents and use them to obtain more information about the complexity of multi-agent systems. In this paper, we propose a novel influence-seeking exploration (ISE) scheme, which encourages agents to preferably explore action spaces significantly influenced by others and thus helps in speeding up the learning curve. To measure the influence of other agents in action selection, we use the variance of joint action-values with different action sets of agents that obtained by an estimation technique to lessen computation overhead. To this end, we first present an analytical approach inspired by the concept of approximated variance propagation and then apply it to an exploration scheme. We evaluate the proposed exploration method on a set of StarCraft II micromanagement as well as modified predator-prey tasks. Compared to state-of-the-art methods, the proposed method achieved performance improvements of 10% in StarCraft II micromanagement and 50% in modified predator-prey tasks approximately.
18 October 2023 in IEEE Access
In this study, we present an end-to-end framework for active speaker detection to achieve robust performance in challenging scenarios with multiple speakers. In contrast to recent approaches, which rely heavily on the visual relational context between all speakers in a video frame, we propose collaboratively learning multimodal representations based on the audio and visual signals of a single candidate. Firstly, we propose a spatio-positional encoder to effectively address the problem of false detections caused by indiscernible faces in a video frame. Secondly, we present an efficient multimodal approach that models the long-term temporal contextual interactions between audio and visual modalities. Through extensive experiments on the AVA-ActiveSpeaker dataset, we demonstrate that our framework notably outperforms recent state-of-the-art approaches under challenging multi-speaker settings. Additionally, the proposed framework significantly improves the robustness against auditory and visual noise interference without relying on pre-trained networks or hand-crafted training strategies.