生成式AI GUI测试的目标.pdf

1、Gen AI GIU testing is theDr.Emil AlgrothEmil.AlegrothBTH.se2Graphical User Interface TestingY pxX pxY pxDistanceFont2Blekinge institute of technologyChallenges with GUI Testing3 Resources(Trillions of dollars/year)Effective/Efficient?Knowledge Domain Technical Non-functionalGenerative AI(Gen AI)Larg

2、e-language models(GPT,BERT,PaLM)Text/language Generative Pre-trained Transformers(GPT)(GPT-X,DALL-E)Text/Images Other(Bark and Chirp)Music Gen AI will add trillions of dollars each yearM.Chui et al.,“The economic potential of generative AI”,McKinsey and Company,2023.4GenAI Collaboration Demo5 Write

3、a song that incorporates the following keywords:Software Engineering Re-thought,SERT,Conference,Exciting,Academic innovation,In Stockholm 20-21 of November 2024.Tony held a fun keynote.Testing with GenAI LLMs improve efficiency Test data generation Test generation Synthesis Challenges Context Errors

4、/hallucinations Competence LLMs might not(EVER)be good enough6Probability?78I,for one,welcome our AI overlords!Even if their current objective is just to predict the next token of a text9Prompts and model engineering Chain/tree of thoughts Break down the task into smaller steps Zero to Multi-shot pr

5、ompting No examples Contextual prompting and tuning Include relevant/iterate context information Model adjustments Temperature Top-K sampling Token frequency Etc.9Agents Contextualization Agents have roles Divide-and-conquer Abstract Agents Solve task X Contextual Agents Act as Agent of Type X10Basi

6、c Agent11Basic Agent12Basic Agent13Why agents?Multiple perspectives Coverage User needs User experience Knowledge New types of testing“Reduce training”14T.A.R.G.E.T.Vinnova funded project 10 MSEK over 3 years Goal Innovate new knowledge and technology to support easy-to-use,end-to-end AI,for efficie