{
  "docId": "019dd923-5e88-73ef-bd5c-f6fdd8f57895",
  "docSlug": "3a892393f72c1ff5",
  "documentTitle": "2019 Air Street Capital The State of AI Report 2019",
  "authorId": "AirStreetCapital",
  "authorName": "Nathan Benaich and Ian Hogarth",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 18,
  "pageCount": 136,
  "prevPage": 17,
  "nextPage": 19,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "overcrowded",
  "nDataPoints": 30,
  "notes": "The slide compares PlaNet's performance against established RL benchmarks (A3C, D4PG) using both a summary table and performance charts.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "data_table",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5c-f6fdd8f57895/18",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5c-f6fdd8f57895",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5c-f6fdd8f57895.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5c-f6fdd8f57895#slide-18",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Within 500 episodes, PlaNet outperforms A3C trained from 100,000 episodes, on six simulator control tasks.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ab-72d9-bbc2-913d96851487",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.4,
        "x": 0.55,
        "y": 0.48
      },
      "kind": "chart",
      "text": "Performance charts for 6 control tasks showing PlaNet vs baselines over 2000 episodes.",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "71aff0b9-f4eb-42dc-b05a-25b6dc76980a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "list",
      "text": "An RL agent (PlaNet) learns a model of the environment's dynamics from images, selects actions through fast online planning by accurately predicting the rewards ahead for multiple time steps.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1ec12a72-ffd4-405e-8cad-b025e8bdbdae",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.45,
        "x": 0.02,
        "y": 0.47
      },
      "kind": "list",
      "text": "After 2,000 episodes, PlaNet achieves similar performance to D4PG, which is trained from images for 100,000 episodes.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1edf718e-b045-42ee-aa3f-db74ce25d15b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.95,
        "x": 0.02,
        "y": 0.31
      },
      "kind": "list",
      "text": "This results in 50x less simulated environmental interactions and similar computation time compared to the state-of-the-art A3C and D4PG algorithms. Within 500 episodes, PlaNet outperforms A3C trained from 100,000 episodes, on six simulator control tasks. This is a significant improvement in data efficiency.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "461f163c-ff36-49f8-9ed1-5b936d41e91b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Data efficiency gain: 180x",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ab-72d9-bbc2-968f06b06455",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.5,
        "x": 0.02,
        "y": 0.65
      },
      "kind": "table",
      "text": "Comparison table of RL methods (A3C, D4PG, PlaNet, CEM) across 6 tasks with modality and episode counts.",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6c3d798f-10d7-46cd-b045-9cce4b65dc60",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.13
      },
      "kind": "title",
      "text": "What's next in RL: Learning dynamics models for online planning",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1cc0cfaa-ff7a-4950-bafb-e778850caa6d",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-0fd5-7148-8ecc-c2c178cb66c5",
      "evidence": "Title states 180x sample-efficiency over A3C.",
      "confidence": 80
    },
    {
      "name": "Chart Selection Guide",
      "slug": "chart-selection-guide",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "019dd95a-0fd5-7148-8ecc-c7259a03560c",
      "evidence": "Comparative learning-curve chart.",
      "confidence": 60
    },
    {
      "name": "Chart Selection Guide",
      "slug": "chart-selection-guide",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "110dd414-0c8c-4ae3-923e-a040e24c6cf6",
      "evidence": "Performance charts for 6 control tasks showing PlaNet vs baselines over 2000 episodes.",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 92,
      "from": 5,
      "beatId": "019dd95a-0682-776c-8e34-ce57e5bbe574",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Definitions through Research/Talent/Industry — fact-dense case studies and benchmarks.",
      "position": 1,
      "confidence": 70,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 92,
      "from": 9,
      "beatId": "019dd95a-0682-776c-8e34-ddbccb391af1",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Unknown",
      "beatSlug": "voyage-return-the-unknown",
      "evidence": "Research/Talent/Industry frontiers explored.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Complication",
      "parentBeatSlug": "complication"
    }
  ],
  "loops": [
    {
      "to": 19,
      "from": 16,
      "name": "Zoom In",
      "slug": "06-zoom-in",
      "bestFor": "Technical deep-dives, case studies, detailed analysis",
      "matchId": "019dd95a-07fe-70ce-8d3c-9eda9346282a",
      "evidence": "Series titled 'What's next in RL: ...' each zooming on a sub-area.",
      "position": 3,
      "objective": "Drill into specific RL research frontiers (sim-to-real, planning, production)",
      "structure": "The Big Picture -> Key Area of Focus -> Specific Detail -> Implication",
      "confidence": 65,
      "description": "Start broad, then progressively focus on specific details that prove your point"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}