International Journal of Smart Computing and Artificial Intelligence

Improving Abstractive Summarization by Transfer Learning with Adaptive Document Selection

Masato Shirai, Kei Wakabayashi — Thu, 19 Oct 2023 13:32:22 +0000

Abstractive document summarization based on neural networks is a promising approach to generate a flexible summary but requires a large amount of training data.
While transfer learning can address this issue, there is a potential concern about the negative transfer effect that deteriorates the performance when we use training documents irrelevant to the target domain, which has not been explicitly explored in document summarization tasks.
In this paper, we propose a method that selects training documents from the source domain that are expected to be useful for the target summarization.
The proposed method is based on the similarity of word distributions between each source document and a set of target documents.
We further propose an adaptive approach that builds a custom-made summarization model for each test document by selecting source documents similar to the test document.
In the experiment, we confirmed that the negative transfer actually happens also in the document summarization tasks.
Additionally, we show that the proposed method effectively avoids the negative transfer issue and improves summarization performance.

An Extension of Particle Swarm Optimization to Identify Multiple Peaks using Re-diversification in Static and Dynamic Environments

Stephen Raharja, Toshiharu Sugawara — Thu, 19 Oct 2023 15:05:33 +0000

We propose an extension of the particle swarm optimization (PSO) algorithm for each particle to store multiple global optima internally for identifying multiple (top-k) peaks in static and dynamic environments. We then applied this technique to search and rescue problems of rescuing potential survivors urgently in life-threatening disaster scenarios. With the rapid development of robotics and
computer technology, aerial drones can be programmed to implement search algorithms that locate potential survivors and relay their positions to rescue teams. We model an environment of a disaster area with potential survivors using randomized
bivariate normal distributions. We extended the Clerk-Kennedy PSO algorithm as top-k PSO by considering individual drones as particles, where each particle remembers a set of global optima to identify the top-k peaks. By comparing several other
algorithms, including the canonical PSO, Clerk-Kennedy PSO, and NichePSO, we evaluated our proposed algorithm in static and dynamic environments. The experimental results show that the proposed algorithm was able to identify the top-k
peaks (optima) with a higher success rate than the baseline methods, although the rate gradually decreased with increasing movement speed of the peaks in dynamic environments.

Improving Multi-Agent Reinforcement Learning for Beer Game by Reward Design Based on Payment Mechanism

Masaaki Hori, Toshihiro Matsui — Fri, 03 Nov 2023 09:27:13 +0000

Supply chain management aims to maximize profits among supply chain partners by managing the flow of information and products. Multiagent reinforcement learning in artificial intelligence research fields has been applied to supply chain management. The beer game is an example problem in supply chain management and has also been studied as a cooperation problem in multiagent systems. In the previous study, a solution method SRDQN that is based on deep reinforcement learning and reward shaping has been applied to the beer game. By introducing a single reinforcement learning agent with SRDQN as a participant in the beer game, the cost of beer inventory was reduced. However, the previous study has not addressed the case of multiagent reinforcement learning due to the difficulties in cooperation among agents. To address the multiagent cases, we apply a reward shaping technique RDPM based on mechanism design to SRDQN and improve cooperative policies in multiagent reinforcement learning. Furthermore, we propose two reward design methods with modifications to the state value function designs in RDPM to address various consumer demands for beers in the supply chain. And then we empirically evaluate the effectiveness of the proposed approaches.

Asset Allocation Method Based on Sentiment Signals and Causal Information using Multi-asset Classes

Rei Taguchi, Hiroki Sakaji, Kiyoshi Izumi, Yuri Murayama — Thu, 14 Dec 2023 09:15:32 +0000

In this study, we demonstrate the usefulness of financial text for asset allocation with multi-asset classes, including stocks and bonds, by creating polarity indexes for several types of financial news through natural language processing. We performed clustering using a change-point detection algorithm on the created polarity indexes. We also constructed a multi-asset portfolio using an optimization algorithm and rebalanced it based on the detected change points. The results show that the proposed asset allocation method performed better than the comparison method, suggesting that polarity indexes can be useful in constructing asset allocation methods with multi-asset classes.