電腦科學與資訊工程

強化學習為當前AI領域的熱門話題，其特點是在環境的獎勵與懲罰下，進行學習。強化學習雖然較為困難，但其成功的項目都非常有名，其中最著名的例子有: AlphoGo、AlphaZero等等。 深度強化學習(DRL)是深度學習與強化學習的結合體，本專題透過DRL實現近端優化策略演算法，來使BipedalWalker環境中的二足模型學會行走，並調適超參數與神經網路來讓模型訓練擁有更好的結果。 經過實驗後發現，適當的降低獎勵折扣衰減率能有效的提升學習速度以及學習上限，同時可以避免分數落差過大導致的Dead relu問題。最終的結果能讓平均分數達到302分，成功達成了BipedalWalker環境要求(平均分數>=300分)。 為了使智能體擁有更好的探索能力，本專題加入了ICM(Intrinsic Curiosity Module)，成功提升了最終的平均分數至316分，將不摔倒的機率提升至99%，最高分數則到了320分，使得雙足模型能以更快的速度向前移動並保持穩定。

密碼為我們生活中常見的一項工具，它可以為我們保護我們的隱私，以免受到別人侵犯，但也同時間出現許多問題，例如：若密碼太過單純會使密碼的功能失去意義，讓其他人可以輕易存取使用者的資料。 在臺灣，注音文密碼是十分常用的密碼，因為它乍看像亂碼，其實有一定的規律存在，而這些密碼，卻容易被判斷為安全的密碼，因此，我們希望可以將這個問題改善。 在研究中，我們先研究密碼強弱，再探討注音文密碼中在一般密碼中的比例，最後達成我們希望的目標—寫出一個可以偵測加入注音文策略的程式。在研究中，我們亦討論各個密碼演算法的優點及缺點，以找出可以最準確判斷的程式，並在研究的最後，提出可以將這套方法擴充至其他語言或是輸入法的可能性。

The Information Technology (IT) area has shown great growth in recent years, even with the economic recession that 巴西 has been through and the impact of the coronavirus pandemic. It is estimated that by 2024 the area will have a deficit of more than 290 thousand professionals. However, companies still face other difficulties in hiring, especially people who are looking for their first job in the Information Technology area. Most part of these difficulties are lack of qualified manpower and high prerequisites to fill internship or junior positions. As a result, the objective of this project is: to develop a platform that connects people who seek guidance, improvement or professional relocation in the Information Technology area with professionals that already have the experience they are seeking. The first step was a research and analysis of similar platforms in the market, whose proposal involves mentoring or professional connections, and it concluded that there are no services that fully meet the project’s proposal. In the second step, a research was done about mobile development, highlighting Flutter and Firebase platform. The third step defined the application’s features, such as suggestion of users and mentors, search for users, become a mentor, private chat, video calls, Portuguese and English languages, light and dark themes and profile customization. The suggestion of users and mentors is done by a match with the registered users, relating their areas of work (where the user has experience) and the areas of interest of each one. For the coding of the project, Flutter and Firebase technologies were used. To design the app, it followed Material Design specifications. For testing and distribution, the app was published on Play Store, Google’s Android application platform. The tests were performed by both the researcher and a selected group of users to verify if the functionalities were in accordance to what was defined in the beginning of the project. Perceiving the correct functioning of the application, the project achieved the proposed objective. In addition, it expanded its reach area, because it is possible to find users and mentors from any other area of the market.

本研究旨在設計一個使用模式，以不切換中、英文輸入法打字的原則之下，能夠完整的自動辨識出一個包含中文（注音、嘸蝦米、倉頡）與英文完整句子。經實測結果，正確率達到94.23%以上。

Stochastic gradient descent (SGD) is one of the preferred online optimization algorithms. However, one of its major drawbacks is its predisposition to forgetting previous data when optimizing through a data stream, also known as catastrophic interference. In this project, we attempt to mitigate this drawback by proposing a new low-cost approach which incorporates backward queried images with SGD during online training. Under this new approach, we propose that for every new training sample through the data stream, the neural network is optimized using the corresponding backward queried image from the initial dataset. After compiling the accuracy of the proposed method and SGD under a data-stream of 50,000 training cases with 10,000 test cases and comparing our algorithm to SGD, we see substantial improvements in the performance of the neural network with two different MNIST datasets (Fashion and Kuzushiji), classifying the MNIST datasets at a high accuracy for the mean, minimum, lower quartile, median, and upper quartile, while maintaining lower standard deviation in performance, demonstrating that our proposed algorithm can be a potential alternative to online SGD.

本研究主要考慮在盡量保留可讀性的情況下，找出將 Imperative Programming 程式碼對應的 Functional Programming 的程式碼並證明。 結果如下： 一、if statement 等價於由 ifte 函數所構成的程式碼，其中函數ifte定義在本文內 二、某些 for-loop statement 等價於由 foldl 函數所構成的程式碼 三、某些 for-loop statement 等價於由 map 函數所構成的程式碼

在資訊爆炸的時代，效率閱讀、整理資料的能力越趨重要。身為高中生，學習時的閱讀量龐大，還須另外自己挑選重點句，重新整理筆記。因此我想如果可以讓電腦自動摘取文章的重點，就能幫助學生效率學習。 大多數現存的自動摘要研究適用於英文文本，本研究利用演算法抓取中文文章的摘要，使學生可以真正實用該演算法於日常學習當中。除此之外，此研究比較了不同方法摘要的準確率以及優缺點。

本研究以AI虛擬試衣系統(Virtual Try-on)為主題，透過深度學習技術，並結合幾何匹配模型，開發出試衣系統，可將使用者上傳的照片，模擬成穿著新衣的模樣。 首先，以深度學習模型將人物原始圖片取出骨架節點，並生成人體遮罩以及保留人物頭部，再結合以上三種資訊合成為高維特徵圖。接著將目標替換衣物生成出依照人體姿態扭曲後的衣物圖片。最後於Virtual Try-on模型中將人體高維特徵圖與扭曲衣物作為輸入，並經過深度學習網路合成出穿著目標衣物之人體圖像。本研究結果發現，人物站姿單純，且雙手緊貼身側，以及拍攝角度為正面、衣服款式為短袖、背景色彩對比度較高與衣服圖案單純的原始圖片，可得到較好的合成結果。

全球貧血人口普遍，然許多人並不了解自身是否罹患貧血；長期患有貧血的病人，亦需定期抽血檢驗追蹤是否有貧血惡化達到需接受輸血的程度。研究顯示，結膜之顏色與貧血有絕對關係，結膜越白則貧血越嚴重，醫師也常使用結膜顏色推測是否有貧血情形。若能設計手機軟體自動分割結膜影像並分析其顏色，將有機會推測受試者是否罹患貧血。本研究收集22位無貧血者及8位貧血病人，並獲得其近期血紅素數值。以手機取得受試者之眼睛影像後，成功設計程式以深度學習完成結膜自動影像分割，對於分割影像以面積大小進行後期處理後，依其取得下眼瞼結膜之三原色平均，再利用kNN與SVM演算法判斷預測出該受試者是否具有貧血之症狀。本研究主要分為兩階段，其一為進行下眼瞼結膜分割模型訓練；其二為製作有無貧血之判斷模型。整合上述眼瞼分割模型(IoU=89.8%±0.02%)與貧血判斷模型(SVM以polynomial核函數測出 準確值93.3%±24.3%)後，可得貧血診斷準確率為80%。此結果代表AI技術有機會透過結膜影像，判斷被拍攝者是否有貧血情形，未來若能增加研究人數，將可設計網頁版或手機APP加以推測血紅素值，供大眾居家篩檢。

Two years ago, I attended a robot contest, in which one of the missions required the robot to follow the pedestrian to complete the task. At that time, I used their demo program to complete the task. Not long after, I found two main issues: 1. The program follows the closest point read by the depth camera, which if I walk close to a wall next to, the robot may likely ‘follow’ the wall. 2. Not to mention if another pedestrian crosses between the robot and the target. Regarding these two issues, I decided to improve it. We’ve designed a procedure of using YOLO Object Detection and Person re-identification to re-identify the target for continuous following.