[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-lexfridman--deeptraffic":3,"tool-lexfridman--deeptraffic":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",138956,2,"2026-04-05T11:33:21",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":78,"owner_company":79,"owner_location":80,"owner_email":81,"owner_twitter":75,"owner_website":82,"owner_url":83,"languages":84,"stars":89,"forks":90,"last_commit_at":91,"license":79,"difficulty_score":92,"env_os":93,"env_gpu":93,"env_ram":93,"env_deps":94,"category_tags":97,"github_topics":98,"view_count":10,"oss_zip_url":81,"oss_zip_packed_at":81,"status":16,"created_at":107,"updated_at":108,"faqs":109,"releases":120},1014,"lexfridman\u002Fdeeptraffic","deeptraffic","DeepTraffic is a deep reinforcement learning competition, part of the MIT Deep Learning series.","DeepTraffic 是 MIT 深度学习系列课程推出的一个深度强化学习竞赛平台，旨在通过游戏化的方式解决密集交通场景下的车辆导航难题。参与者需要设计神经网络来控制一辆或多辆车，在拥堵的高速公路上以最快的速度安全行驶，同时避免碰撞。\n\n这个平台直面现代交通的核心痛点——拥堵。仅在美国，每年就有 69 亿小时被浪费在交通堵塞中。DeepTraffic 为自动驾驶运动规划算法提供了一个理想的实验场，让研究者探索如何减少幽灵堵车、提升道路通行效率。\n\nDeepTraffic 适合从初学者到专家各类水平的研究人员和开发者。新手可以通过修改预设的神经网络代码快速上手，而资深研究者则能深入调优模型，在排行榜上与全球参与者一较高下。平台提供实时可视化、网络激活状态监控和在线训练测试环境，让算法调试直观易懂。\n\n其技术亮点在于将众包超参数调优与多智能体深度强化学习相结合，让数千名参与者的集体智慧共同推动复杂交通场景下的 AI 决策能力边界。所有提交方案均可生成可视化回放，便于分享和分析。","\n\n# DeepTraffic: MIT Deep Reinforcement Learning Competition\n\n[DeepTraffic](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic) - [Visualization](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-visualization) - [Leaderboard](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-leaderboard) - [Documentation](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation) - [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805) - [MIT Deep Learning](https:\u002F\u002Fdeeplearning.mit.edu\u002F) [ [GitHub](https:\u002F\u002Fgithub.com\u002Flexfridman\u002Fmit-deep-learning) | [Website](https:\u002F\u002Fdeeplearning.mit.edu\u002F) ]\n\nDeepTraffic is a deep reinforcement learning competition hosted as part of the [MIT Deep Learning](https:\u002F\u002Fdeeplearning.mit.edu) courses. The goal is to create a neural network that drives a vehicle (or multiple vehicles) as fast as possible through dense highway traffic. Top 10 submissions are listed on the  [leaderboard](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-leaderboard\u002F) and you'll be able to [visualize](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-visualization\u002F) your submission in the following way:\n\n![DeepTraffic visualization](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Flexfridman_deeptraffic_readme_fe2d2d619aa0.gif)\n\nIf you find the work useful in your research, please cite the [DeepTraffic paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805):\n\n```bibtex\n@inproceedings{fridman2018deeptraffic,\nauthor = {Lex Fridman and Jack Terwilliger and Benedikt Jenik},\ntitle = {DeepTraffic: Crowdsourced Hyperparameter Tuning of Deep Reinforcement Learning Systems for Multi-Agent Dense Traffic Navigation},\nbooktitle = {Neural Information Processing Systems (NIPS 2018) Deep Reinforcement Learning Workshop}\nyear = {2018},\nurl = {http:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805},\ndoi = {10.5281\u002Fzenodo.2530457}\narchivePrefix = {arXiv},\n}\n```\n\nTo get started right away, this repository provides a code snippet to insert into the code box on the [DeepTraffic site](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic\u002F). We'll add additional agents  as the course progresses:\n\n**network_basic.js**: A basic network that achieves ~66.8mph.\n\nAnd now let's return to the problem of traffic:\n\n## Problem Statement: Traffic is Terrible\n\n> \"Americans will put up with anything provided it doesn’t block traffic.\" - Dan Rather \n\n> \"Traffic is soul-destroying.\" - Elon Musk\n\nIn the U.S. alone, we spend 6.9 billion hours sitting in traffic each year [1] — roughly 10,000 human lifetimes [2]. Autonomous vehicles will be able to alleviate part (but not all) of the problem. Already, they show promise in reducing phantom traffic jams [3,4].\n\nWe’ve designed DeepTraffic to let people (from beginners to experts) explore the design of motion planning algorithms for autonomous vehicles and to inspire the next generation of traffic engineering. We thank the thousands of competitors who have submitted solutions and are actively participating.\n\n## DeepTraffic Layout\n\n\u003Cimg src=\"https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2018\u002F01\u002FScreenshot-from-2018-01-08-17-45-42.png\" \nalt=\"DeepTraffic\" width=\"600\" \u002F>\n\nThe game page consists of four different areas:\n\n- On the left, you can find a real time simulation of the road with different display options.\n\n- On the upper half of the page, you can find (1) a coding area where you can change the design of the neural network which controls the agents and (2) some buttons for applying your changes, saving\u002Floading, and making a submission.\n\n- Below the coding area, you can find (1) a graph showing a moving average of the center red car’s reward, (2) a visualization of the neural network activations, and (3) buttons for training and testing your network.\n\n- Between the simulated roadway and the graphs, you can find the current image of you vehicle and some options to customize it and create a visualization of your best submission.\n\nThe simulation area shows some basic information like the current speed of the car and the number of cars that have been passed since you opened the site. It also allows you to change the way the simulation is displayed.\n![Display selection](http:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002Fnormal.gif)\n\n## DeepTraffic Simulation & Game\n\nIn short, DeepTraffic is a game in which you (a competitor) design your own motion planning algorithm in order to drive a vehicle as fast as possible through dense traffic.\n\nYour algorithm will operate on a 7 lane highway. There are 20 vehicles on the road. Your algorithm controls some vehicles. The game controls the others.\n\nEach autonomous agent runs a copy of your algorithm. Every 30 frames, your algorithm selects 1 of 5 actions:\n\n1. accelerate\n1. decelerate\n1. change into the left lane\n1. change in to the right lane\n1. do nothing, i.e. maintain speed in present lane.\n\nYour algorithm will receive, as input, an occupancy grid, representing the free space around the agent. The value of unoccupied cells is set to 80mph. The value of occupied cells is set to the speed of the occupying vehicle. For example here's an occupancy grid `(lanesSide = 1; patchesAhead=10)`:\n\n![learning input](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002FScreen-Shot-2017-01-03-at-16.06.29-151x300.png)\n\nThere are a few quirks to DeepTraffic’s dynamics:\n\n### Safety System\n\nEach vehicle has a safety system which prevents it from colliding with other vehicles. This has 2 implications for how you will design your algorithm. First, your algorithm does not need to consider collision avoidance. Second, your path will be overridden when the safety system is activated.\n\nFor example, here, the red car cannot accelerate or change into the right lane, because the collision avoidance system has detected vehicles in the way:\n\n![safety system](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002FScreen-Shot-2017-01-03-at-16.28.32-152x300.png)\n\nA vehicle that is 4 cells behind another will immediately slow to match the lead vehicle, regardless of what its algorithm tries to do. (see diagram above)\n\nA vehicle driving beside another will be unable to change into its neighbor’s lane until there is a sufficient gap, regardless of what its algorithm tries to do. (see diagram above)\n\n### Multiple Agents\n\nIn version 2.0, the current version, you have the option to deploy a copy of your algorithm on 11 vehicles. You algorithm won’t do multi-agent planning, rather each vehicle makes a greedy choice. The challenge is to design an algorithm which does not get in its own way when controlling several vehicles.\n\n### Where the Highway Ends\n\nDeepTraffic follows just one of the vehicles (the ego vehicle), so you’ll notice some of the vehicles fall off the highway when they drive slower or faster than the ego vehicle. What happens to these vehicles?\n\nWhen vehicles fall off the road, they are replaced by new vehicles on the opposite end of the highway. When a vehicle is replaced, its speed & lane is chosen randomly.\n\n## Hyperparameters\n\nTo do well in DeepTraffic using DQN, you’ll have to choose good hyperparameters. This can be tricky because (1) the full hyperparameter space is rather large and (2) the bigger your network gets, the longer it takes to train which means you’ll explore less of the hyper-parameter space. Therefore, it helps to understand how changing the hyper-parameters will change performance prior to training.\n\n![parameters](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2019\u002F01\u002Fdeeptraffic_parameters.png)\n\n## Results\n\n### Progress\n\nThe plot below shows how the competition progressed over time:\n\n![progress](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2019\u002F01\u002Fscore_curve_data.png)\n\n### The Structure of Submissions\n\nBelow is a t-SNE plot, i.e. submissions originally represented in a vector space spanning patchesAhead, patchesBehind, l2_decay, layer_count, gamma, learning_rate, lanesSide, train_iterations are plotted in a 2 dimensional space which preserves the composition of neighboring points. The color of each dot corresponds to submissions score. An interesting feature of this plot is that several clusters emerge — competitors stumbled upon similar solutions.\n\n![tsne](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2018\u002F06\u002Ftsne-scatter.png)\n\n## Help and Documentation\n\nSee [Documentation page](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation\u002F) for more details and hints and how to submit to the competition.\n\n## Team\n\n- [Lex Fridman](https:\u002F\u002Flexfridman.com\u002F) ([Twitter](https:\u002F\u002Ftwitter.com\u002F))\n- [Jack Terwilliger](http:\u002F\u002Fwww.mit.edu\u002F~jterwill\u002F)\n- [Benedikt Jenik](http:\u002F\u002Fbjenik.com\u002F)\n\n## References\n\n- [1] https:\u002F\u002Fstatic.tti.tamu.edu\u002Ftti.tamu.edu\u002Fdocuments\u002Fmobility-scorecard-2015.pdf\n- [2] (6.9 * 1000000000) \u002F (75 * 365 * 24)\n- [3] Horn, Berthold KP. \"Suppressing traffic flow instabilities.\" Intelligent Transportation Systems-(ITSC), 2013 16th International IEEE Conference on. IEEE, 2013. https:\u002F\u002Fpeople.csail.mit.edu\u002Fbkph\u002Farticles\u002FSuppressing_Traffic_Flow%20Instabilities_IEEE_ITS_2013.pdf\n- [4] Stern, Raphael E., et al. \"Dissipation of stop-and-go waves via control of autonomous vehicles: Field experiments.\" Transportation Research Part C: Emerging Technologies 89 (2018): 205-221. https:\u002F\u002Farxiv.org\u002Fpdf\u002F1705.01693.pdf\n- [5] https:\u002F\u002Flink.springer.com\u002Fcontent\u002Fpdf\u002F10.1007\u002FBF00992698.pdf\n","# DeepTraffic: MIT 深度强化学习竞赛\n\n[DeepTraffic](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic) - [可视化](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-visualization) - [排行榜](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-leaderboard) - [文档](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation) - [论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805) - [MIT 深度学习](https:\u002F\u002Fdeeplearning.mit.edu\u002F) [ [GitHub](https:\u002F\u002Fgithub.com\u002Flexfridman\u002Fmit-deep-learning) | [网站](https:\u002F\u002Fdeeplearning.mit.edu\u002F) ]\n\nDeepTraffic 是作为 [MIT 深度学习](https:\u002F\u002Fdeeplearning.mit.edu) 课程一部分举办的深度强化学习（deep reinforcement learning）竞赛。目标是创建一个神经网络（neural network），使其能够在密集的高速公路交通中尽可能快地驾驶车辆（或多辆车辆）。排行榜前 10 名的提交将显示在 [排行榜](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-leaderboard\u002F) 上，您将能够以以下方式 [可视化](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-visualization\u002F) 您的提交：\n\n![DeepTraffic visualization](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Flexfridman_deeptraffic_readme_fe2d2d619aa0.gif)\n\n如果您在研究中发现该工作有用，请引用 [DeepTraffic 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805)：\n\n```bibtex\n@inproceedings{fridman2018deeptraffic,\nauthor = {Lex Fridman and Jack Terwilliger and Benedikt Jenik},\ntitle = {DeepTraffic: Crowdsourced Hyperparameter Tuning of Deep Reinforcement Learning Systems for Multi-Agent Dense Traffic Navigation},\nbooktitle = {Neural Information Processing Systems (NIPS 2018) Deep Reinforcement Learning Workshop}\nyear = {2018},\nurl = {http:\u002F\u002Farxiv.org\u002Fabs\u002F1801.02805},\ndoi = {10.5281\u002Fzenodo.2530457}\narchivePrefix = {arXiv},\n}\n```\n\n要立即开始，本仓库提供了一段代码片段，可插入到 [DeepTraffic 网站](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic\u002F) 的代码框中。随着课程的进展，我们将添加更多智能体（agents）：\n\n**network_basic.js**：一个基础网络，可达到约 66.8 英里\u002F小时的速度。\n\n现在让我们回到交通问题：\n\n## 问题陈述：交通状况糟糕\n\n> \"只要不影响交通，美国人什么都能忍受。\" —— Dan Rather \n\n> \"交通让人精神崩溃。\" —— Elon Musk\n\n仅在美国，我们每年就花费 69 亿小时坐在交通拥堵中 [1] —— 大约相当于 10,000 个人一生的时间 [2]。自动驾驶汽车（autonomous vehicles）将能够缓解部分（但不是全部）问题。它们已经在减少幽灵交通拥堵（phantom traffic jams）方面显示出潜力 [3,4]。\n\n我们设计 DeepTraffic 的目的是让人们（从初学者到专家）探索自动驾驶车辆的运动规划算法（motion planning algorithms）设计，并激励下一代交通工程的发展。我们感谢数千名提交解决方案并积极参与的参赛者。\n\n## DeepTraffic 布局\n\n\u003Cimg src=\"https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2018\u002F01\u002FScreenshot-from-2018-01-08-17-45-42.png\" \nalt=\"DeepTraffic\" width=\"600\" \u002F>\n\n游戏页面由四个不同的区域组成：\n\n- 在左侧，您可以找到道路的实时模拟，具有不同的显示选项。\n\n- 在页面的上半部分，您可以找到（1）一个代码编辑区域（coding area），您可以在其中更改控制智能体（agents）的神经网络设计，以及（2）一些用于应用更改、保存\u002F加载和提交结果的按钮。\n\n- 在代码编辑区域下方，您可以找到（1）显示中央红色车辆奖励（reward）移动平均值的图表，（2）神经网络激活（activations）的可视化，以及（3）用于训练和测试网络的按钮。\n\n- 在模拟道路和图表之间，您可以找到您车辆的当前图像以及一些自定义选项，用于创建您最佳提交的可视化效果。\n\n模拟区域显示一些基本信息，如汽车的当前速度以及自您打开网站以来已超越的车辆数量。它还允许您更改模拟的显示方式。\n![Display selection](http:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002Fnormal.gif)\n\n## DeepTraffic 模拟与游戏\n\n简而言之，DeepTraffic 是一个游戏，在其中您（参赛者）设计自己的运动规划算法（motion planning algorithm），以便在密集的交通中尽可能快地驾驶车辆。\n\n您的算法将在一条 7 车道的高速公路上运行。道路上有 20 辆车。您的算法控制其中一些车辆，游戏控制其他车辆。\n\n每个自主智能体（autonomous agent）都运行您的算法的一个副本。每 30 帧，您的算法从 5 个动作中选择 1 个：\n\n1. 加速（accelerate）\n1. 减速（decelerate）\n1. 变换到左侧车道\n1. 变换到右侧车道\n1. 什么都不做（do nothing），即在当前车道保持速度。\n\n您的算法将接收一个占用网格（occupancy grid）作为输入，表示智能体周围的自由空间。未被占用的单元格值设置为 80 英里\u002F小时。被占用的单元格值设置为占用车辆的速度。例如，这是一个占用网格（occupancy grid）`(lanesSide = 1; patchesAhead = 10)`：\n\n![learning input](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002FScreen-Shot-2017-01-03-at-16.06.29-151x300.png)\n\nDeepTraffic 的动力学有一些特殊之处：\n\n### 安全系统\n\n每辆车都有一个防止其与其他车辆碰撞的安全系统（safety system）。这对您设计算法有两个影响。首先，您的算法不需要考虑碰撞避免（collision avoidance）。其次，当安全系统被激活时，您的路径将被覆盖。\n\n例如，在这里，红色汽车无法加速或变换到右侧车道，因为碰撞避免系统（collision avoidance system）已检测到前方有车辆：\n\n![safety system](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2017\u002F01\u002FScreen-Shot-2017-01-03-at-16.28.32-152x300.png)\n\n一辆位于另一辆车后方 4 个单元格的车辆将立即减速以匹配前方车辆的速度，无论其算法试图做什么。（见上图）\n\n一辆在另一辆车旁边行驶的车辆将无法变换到邻车的车道，直到有足够的间隙，无论其算法试图做什么。（见上图）\n\n### 多智能体\n\n在 2.0 版本（当前版本）中，您可以选择在 11 辆车上部署您的算法的一个副本。您的算法不会进行多智能体规划（multi-agent planning），而是每辆车做出贪婪选择（greedy choice）。挑战在于设计一个算法，在控制多辆车时不会妨碍自己。\n\n### 高速公路的尽头\n\nDeepTraffic 只跟随其中一辆车（自车，ego vehicle），因此您会注意到当某些车辆比自车开得慢或快时，它们会从高速公路上消失。这些车辆会发生什么？\n\n当车辆从道路上消失时，它们会在高速公路的另一端被新车辆取代。当车辆被替换时，其速度和车道将被随机选择。\n\n## 超参数\n\n要在 DeepTraffic 中使用 DQN（Deep Q-Network，深度 Q 网络）取得好成绩，你必须选择好的超参数。这可能很棘手，因为 (1) 完整的超参数空间相当大，(2) 网络越大，训练所需的时间就越长，这意味着你能探索的超参数空间就越少。因此，在训练前了解改变超参数将如何影响性能是很有帮助的。\n\n![parameters](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2019\u002F01\u002Fdeeptraffic_parameters.png)\n\n## 结果\n\n### 进展\n\n下图展示了比赛随时间推移的进展情况：\n\n![progress](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2019\u002F01\u002Fscore_curve_data.png)\n\n### 提交的结构\n\n下面是一个 t-SNE（t-分布随机邻域嵌入）图。该图将原本在向量空间中的提交（此空间涵盖 patchesAhead、patchesBehind、l2_decay、layer_count、gamma、learning_rate、lanesSide、train_iterations 等维度）映射到二维空间，同时保留邻近点的关系结构。每个点的颜色对应提交的分数。这张图的一个有趣特征是出现了几个聚类——参赛者找到了相似的解决方案。\n\n![tsne](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fwordpress\u002Fwp-content\u002Fuploads\u002F2018\u002F06\u002Ftsne-scatter.png)\n\n## 帮助与文档\n\n查看[文档页面](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation\u002F)以获取更多详情、提示以及如何向比赛提交。\n\n## 团队\n\n- [Lex Fridman](https:\u002F\u002Flexfridman.com\u002F) ([Twitter](https:\u002F\u002Ftwitter.com\u002F))\n- [Jack Terwilliger](http:\u002F\u002Fwww.mit.edu\u002F~jterwill\u002F)\n- [Benedikt Jenik](http:\u002F\u002Fbjenik.com\u002F)\n\n## 参考文献\n\n- [1] https:\u002F\u002Fstatic.tti.tamu.edu\u002Ftti.tamu.edu\u002Fdocuments\u002Fmobility-scorecard-2015.pdf\n- [2] (6.9 * 1000000000) \u002F (75 * 365 * 24)\n- [3] Horn, Berthold KP. \"Suppressing traffic flow instabilities.\" Intelligent Transportation Systems-(ITSC), 2013 16th International IEEE Conference on. IEEE, 2013. https:\u002F\u002Fpeople.csail.mit.edu\u002Fbkph\u002Farticles\u002FSuppressing_Traffic_Flow%20Instabilities_IEEE_ITS_2013.pdf\n- [4] Stern, Raphael E., et al. \"Dissipation of stop-and-go waves via control of autonomous vehicles: Field experiments.\" Transportation Research Part C: Emerging Technologies 89 (2018): 205-221. https:\u002F\u002Farxiv.org\u002Fpdf\u002F1705.01693.pdf\n- [5] https:\u002F\u002Flink.springer.com\u002Fcontent\u002Fpdf\u002F10.1007\u002FBF00992698.pdf","# DeepTraffic 快速上手指南\n\nDeepTraffic 是 MIT 推出的浏览器端深度强化学习竞赛平台，通过设计神经网络控制车辆在密集交通中高速行驶。\n\n## 环境准备\n\n- **现代浏览器**：Chrome\u002FFirefox 最新版\n- **网络连接**：可访问 [https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic\u002F](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic\u002F)\n- **基础知识**：JavaScript 与强化学习基础\n\n## 安装步骤\n\n无需本地安装，完全在线运行。\n\n1. **访问竞赛网站**\n   ```\n   https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic\u002F\n   ```\n\n2. **复制基础代码**\n   打开本仓库的 `network_basic.js` 文件，复制全部代码。\n\n3. **粘贴到编辑器**\n   在网站左侧代码框中粘贴，替换默认代码。\n\n## 基本使用\n\n### 训练与测试\n1. 点击 **Apply** 应用代码更改\n2. 点击 **Train** 开始训练（观察右侧奖励曲线）\n3. 点击 **Test** 评估性能（查看平均速度）\n4. 点击 **Submit to Leaderboard** 提交成绩\n\n### 核心配置\n基础代码已实现约 66.8 mph 速度。关键可调参数：\n- `learning_rate`: 学习率\n- `gamma`: 折扣因子\n- `layers`: 网络层结构\n- `lanesSide`\u002F`patchesAhead`\u002F`patchesBehind`: 感知范围\n\n### 可视化\n训练完成后，可在 [可视化页面](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-visualization\u002F) 查看你的提交记录行驶动画。\n\n**完整参数说明**：[官方文档](https:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation\u002F)","某大学人工智能课程的期末项目中，学生团队需要设计一个能在密集车流中自主导航的驾驶策略，作为强化学习模块的实践考核。\n\n### 没有 deeptraffic 时\n- **环境搭建耗时**：学生需用 Unity 或 CARLA 自行构建高速公路仿真场景，仅配置多车交互逻辑就花费两周时间，核心算法研究被严重压缩\n- **调试过程盲目**：训练时只能看日志数据，无法直观看到车辆决策瞬间的神经网络激活情况，一个参数调错往往要重新训练6小时才能发现\n- **缺乏性能标尺**：团队不知道65mph的测试成绩是好是坏，没有行业基准参考，优化方向全凭感觉，学习动力逐渐消退\n- **协作门槛极高**：非计算机专业学生因环境配置复杂（CUDA、依赖库版本冲突）而难以参与，团队被迫缩减为3名核心成员\n\n### 使用 deeptraffic 后\n- **即开即用的战场**：浏览器打开即可开始，预置的密集交通流和碰撞检测让团队第一天就进入算法设计，将开发周期从3周压缩到5天\n- **实时诊断能力**：训练时同步观察车辆视角、Q值热图和每层神经元激活状态，一次训练就发现卷积层感受野设置过小的问题，调试效率提升80%\n- **全球排行榜驱动**：提交后立刻看到排名，从第200名提升到第47名的过程让团队保持高度专注，MIT官方提供的baseline网络成为明确的超越目标\n- **零门槛团队协作**：前端同学也能在网页上修改奖励函数参数，5人小组各自尝试不同策略并行训练，最终融合方案达到72mph，获得课程最高分\n\ndeeptraffic 将强化学习从\"环境搭建噩梦\"转变为\"算法创新游乐场\"，让学生真正把时间花在探索智能决策的本质上。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Flexfridman_deeptraffic_fe2d2d61.gif","lexfridman","Lex Fridman","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Flexfridman_7c85da30.jpg","AI researcher working on autonomous vehicles, human-robot interaction, and machine learning at MIT and beyond.","MIT","Cambridge, MA",null,"https:\u002F\u002Flexfridman.com","https:\u002F\u002Fgithub.com\u002Flexfridman",[85],{"name":86,"color":87,"percentage":88},"JavaScript","#f1e05a",100,1792,279,"2026-04-02T08:36:30",1,"未说明",{"notes":95,"python":93,"dependencies":96},"DeepTraffic 是基于浏览器的在线竞赛平台，无需本地安装配置。用户直接在网页代码框中编写 JavaScript 来定义神经网络，所有训练与仿真在 MIT 服务器端完成。README 未提供任何本地运行环境的系统要求、依赖库或硬件配置信息。如需离线使用，需自行重构整个系统环境。",[93],[13],[99,100,101,102,103,104,105,106],"deep-learning","machine-learning","deep-reinforcement-learning","mit","self-driving-cars","deep-rl","tensorflow","convnetjs","2026-03-27T02:49:30.150509","2026-04-06T07:12:55.743537",[110,115],{"id":111,"question_zh":112,"answer_zh":113,"source_url":114},4515,"观察值向量的具体含义是什么？为什么被占用的单元格显示的是绝对速度而不是相对速度？","观察值向量的编码规则如下：\n- **0**：表示该单元格位于道路范围外\n- **1**：表示该单元格空闲（未被其他车辆占用）\n- **0 到 1 之间的值**：表示该单元格被车辆占用，具体数值为该车辆的绝对速度除以 2000\n\n关于论文中提到的\"相对速度\"，实际指的是**坐标系的相对性**，而非速度值的相对性。观察值采用自车（ego vehicle）的局部坐标系（相对于自车的位置），而不是全局坐标系。因此，单元格中的速度值是其他车辆的绝对速度，而不是相对于自车的速度差。","https:\u002F\u002Fgithub.com\u002Flexfridman\u002Fdeeptraffic\u002Fissues\u002F7",{"id":116,"question_zh":117,"answer_zh":118,"source_url":119},4516,"我是新手，如何获取 DeepTraffic 的基础代码和入门教程？","可以参考 MIT 官方提供的 DeepTraffic 基础教程文档：\nhttps:\u002F\u002Fselfdrivingcars.mit.edu\u002Fdeeptraffic-documentation\u002F\n\n该文档包含完整的入门指南、基础代码示例和竞赛规则说明，非常适合新手快速上手。","https:\u002F\u002Fgithub.com\u002Flexfridman\u002Fdeeptraffic\u002Fissues\u002F6",[]]