Skip to content

Instantly share code, notes, and snippets.

@jbdamask
Created February 23, 2026 00:10
Show Gist options
  • Select an option

  • Save jbdamask/317b32036eab1280febc08495b20668c to your computer and use it in GitHub Desktop.

Select an option

Save jbdamask/317b32036eab1280febc08495b20668c to your computer and use it in GitHub Desktop.
NowIGetIt: multi.pdf
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>MA-CMAB: Smart Resource Allocation Under Uncertainty</title>
<style>
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700;800;900&family=JetBrains+Mono:wght@400;500&display=swap');
* { margin: 0; padding: 0; box-sizing: border-box; }
:root {
--bg: #0a0a0f;
--surface: #12121a;
--surface2: #1a1a2e;
--accent: #6c63ff;
--accent2: #ff6584;
--accent3: #43e97b;
--accent4: #38f9d7;
--text: #e8e8f0;
--text-dim: #8888a0;
--border: #2a2a3e;
}
html { scroll-behavior: smooth; }
body {
font-family: 'Inter', sans-serif;
background: var(--bg);
color: var(--text);
overflow-x: hidden;
line-height: 1.6;
}
/* Scrollbar */
::-webkit-scrollbar { width: 6px; }
::-webkit-scrollbar-track { background: var(--bg); }
::-webkit-scrollbar-thumb { background: var(--accent); border-radius: 3px; }
/* Canvas backgrounds */
#hero-canvas, #particles-canvas {
position: absolute;
top: 0; left: 0;
width: 100%; height: 100%;
pointer-events: none;
}
/* NAV */
nav {
position: fixed; top: 0; left: 0; right: 0; z-index: 1000;
backdrop-filter: blur(20px);
background: rgba(10,10,15,0.8);
border-bottom: 1px solid var(--border);
padding: 0 2rem;
transition: transform 0.3s;
}
nav .nav-inner {
max-width: 1200px; margin: 0 auto;
display: flex; align-items: center; justify-content: space-between;
height: 64px;
}
nav .logo {
font-weight: 800; font-size: 1.1rem;
background: linear-gradient(135deg, var(--accent), var(--accent2));
-webkit-background-clip: text; -webkit-text-fill-color: transparent;
}
nav .nav-links { display: flex; gap: 2rem; list-style: none; }
nav .nav-links a {
color: var(--text-dim); text-decoration: none; font-size: 0.85rem;
font-weight: 500; transition: color 0.3s; position: relative;
}
nav .nav-links a:hover { color: var(--text); }
nav .nav-links a::after {
content: ''; position: absolute; bottom: -4px; left: 0;
width: 0; height: 2px;
background: var(--accent); transition: width 0.3s;
}
nav .nav-links a:hover::after { width: 100%; }
.hamburger { display: none; flex-direction: column; gap: 5px; cursor: pointer; background: none; border: none; }
.hamburger span { display: block; width: 24px; height: 2px; background: var(--text); transition: 0.3s; }
/* HERO */
.hero {
position: relative; min-height: 100vh;
display: flex; align-items: center; justify-content: center;
text-align: center; padding: 6rem 2rem 4rem;
overflow: hidden;
}
.hero-bg {
position: absolute; inset: 0;
background: radial-gradient(ellipse at 30% 20%, rgba(108,99,255,0.15) 0%, transparent 50%),
radial-gradient(ellipse at 70% 80%, rgba(255,101,132,0.1) 0%, transparent 50%);
}
.hero-content { position: relative; z-index: 2; max-width: 900px; }
.hero-badge {
display: inline-block; padding: 0.4rem 1rem; border-radius: 50px;
font-size: 0.75rem; font-weight: 600; text-transform: uppercase; letter-spacing: 2px;
background: rgba(108,99,255,0.15); color: var(--accent); border: 1px solid rgba(108,99,255,0.3);
margin-bottom: 1.5rem;
}
.hero h1 {
font-size: clamp(2.5rem, 6vw, 4.5rem); font-weight: 900;
line-height: 1.1; margin-bottom: 1.5rem;
}
.hero h1 .gradient-text {
background: linear-gradient(135deg, var(--accent), var(--accent4), var(--accent3));
-webkit-background-clip: text; -webkit-text-fill-color: transparent;
}
.hero p {
font-size: 1.15rem; color: var(--text-dim); max-width: 650px;
margin: 0 auto 2.5rem; line-height: 1.7;
}
.hero-cta {
display: inline-flex; align-items: center; gap: 0.5rem;
padding: 0.9rem 2rem; border-radius: 12px; font-weight: 600;
font-size: 0.95rem; text-decoration: none; color: #fff;
background: linear-gradient(135deg, var(--accent), #8b5cf6);
box-shadow: 0 4px 30px rgba(108,99,255,0.4);
transition: transform 0.3s, box-shadow 0.3s;
}
.hero-cta:hover { transform: translateY(-2px); box-shadow: 0 8px 40px rgba(108,99,255,0.5); }
.hero-stats {
display: flex; justify-content: center; gap: 3rem; margin-top: 3rem;
flex-wrap: wrap;
}
.hero-stat { text-align: center; }
.hero-stat .num {
font-size: 2rem; font-weight: 800;
background: linear-gradient(135deg, var(--accent3), var(--accent4));
-webkit-background-clip: text; -webkit-text-fill-color: transparent;
}
.hero-stat .label { font-size: 0.75rem; color: var(--text-dim); text-transform: uppercase; letter-spacing: 1px; }
/* SECTIONS */
section { padding: 6rem 2rem; position: relative; }
.section-inner { max-width: 1100px; margin: 0 auto; }
.section-label {
font-size: 0.7rem; font-weight: 700; text-transform: uppercase;
letter-spacing: 3px; color: var(--accent); margin-bottom: 0.75rem;
}
.section-title {
font-size: clamp(1.8rem, 4vw, 2.8rem); font-weight: 800;
margin-bottom: 1rem; line-height: 1.2;
}
.section-desc {
font-size: 1.05rem; color: var(--text-dim); max-width: 700px;
margin-bottom: 3rem; line-height: 1.7;
}
/* CARDS */
.card-grid {
display: grid; gap: 1.5rem;
grid-template-columns: repeat(auto-fit, minmax(300px, 1fr));
}
.card {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 16px; padding: 2rem;
transition: transform 0.3s, border-color 0.3s, box-shadow 0.3s;
position: relative; overflow: hidden;
}
.card:hover {
transform: translateY(-4px);
border-color: var(--accent);
box-shadow: 0 8px 40px rgba(108,99,255,0.15);
}
.card-icon {
width: 48px; height: 48px; border-radius: 12px;
display: flex; align-items: center; justify-content: center;
font-size: 1.5rem; margin-bottom: 1.2rem;
}
.card h3 { font-size: 1.15rem; font-weight: 700; margin-bottom: 0.6rem; }
.card p { font-size: 0.9rem; color: var(--text-dim); line-height: 1.6; }
/* Interactive demo area */
.demo-container {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 20px; padding: 2.5rem;
margin-top: 2rem;
}
.demo-header { display: flex; justify-content: space-between; align-items: center; flex-wrap: wrap; gap: 1rem; margin-bottom: 2rem; }
.demo-header h3 { font-size: 1.2rem; font-weight: 700; }
.btn {
padding: 0.6rem 1.5rem; border-radius: 10px; font-weight: 600;
font-size: 0.85rem; border: none; cursor: pointer; transition: all 0.3s;
font-family: 'Inter', sans-serif;
}
.btn-primary { background: var(--accent); color: #fff; }
.btn-primary:hover { background: #7c73ff; transform: translateY(-1px); }
.btn-outline { background: transparent; color: var(--accent); border: 1px solid var(--accent); }
.btn-outline:hover { background: rgba(108,99,255,0.1); }
/* Allocation Viz */
#alloc-viz {
display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
gap: 1.5rem; margin-bottom: 2rem;
}
.agent-box {
background: var(--surface2); border-radius: 14px; padding: 1.5rem;
border: 1px solid var(--border); min-height: 150px;
transition: border-color 0.5s, box-shadow 0.5s;
}
.agent-box.highlight { border-color: var(--accent3); box-shadow: 0 0 20px rgba(67,233,123,0.15); }
.agent-label {
font-size: 0.75rem; font-weight: 700; text-transform: uppercase;
letter-spacing: 1px; margin-bottom: 0.8rem; display: flex;
align-items: center; gap: 0.5rem;
}
.agent-dot { width: 8px; height: 8px; border-radius: 50%; }
.agent-items { display: flex; flex-wrap: wrap; gap: 0.5rem; margin-top: 0.5rem; }
.item-chip {
display: inline-flex; align-items: center; gap: 0.3rem;
padding: 0.35rem 0.75rem; border-radius: 8px;
font-size: 0.75rem; font-weight: 600;
background: rgba(108,99,255,0.12); color: var(--accent);
border: 1px solid rgba(108,99,255,0.2);
animation: chipIn 0.4s ease;
}
@keyframes chipIn {
from { opacity: 0; transform: scale(0.7); }
to { opacity: 1; transform: scale(1); }
}
.welfare-bar-container { margin-top: 1rem; }
.welfare-bar-label { font-size: 0.8rem; color: var(--text-dim); margin-bottom: 0.3rem; }
.welfare-bar-track {
height: 8px; background: var(--surface2); border-radius: 4px; overflow: hidden;
}
.welfare-bar-fill {
height: 100%; border-radius: 4px; transition: width 0.8s ease;
background: linear-gradient(90deg, var(--accent), var(--accent3));
}
.welfare-value {
font-size: 1.5rem; font-weight: 800; margin-top: 0.5rem;
background: linear-gradient(135deg, var(--accent3), var(--accent4));
-webkit-background-clip: text; -webkit-text-fill-color: transparent;
}
/* Regret Chart */
#regret-canvas {
width: 100%; height: 300px; display: block;
background: var(--surface2); border-radius: 14px;
}
/* Algorithm steps */
.algo-timeline {
position: relative; padding-left: 3rem;
}
.algo-timeline::before {
content: ''; position: absolute; left: 14px; top: 0; bottom: 0;
width: 2px; background: var(--border);
}
.algo-step {
position: relative; margin-bottom: 2.5rem; opacity: 0.5;
transition: opacity 0.5s;
}
.algo-step.active { opacity: 1; }
.algo-step::before {
content: ''; position: absolute; left: -2.35rem; top: 4px;
width: 12px; height: 12px; border-radius: 50%;
background: var(--border); border: 2px solid var(--surface);
transition: background 0.5s;
}
.algo-step.active::before { background: var(--accent); box-shadow: 0 0 12px rgba(108,99,255,0.5); }
.algo-step h4 { font-size: 1rem; font-weight: 700; margin-bottom: 0.3rem; }
.algo-step p { font-size: 0.85rem; color: var(--text-dim); }
/* Equation box */
.eq-box {
display: inline-block; padding: 1rem 1.5rem; border-radius: 12px;
background: var(--surface2); border: 1px solid var(--border);
font-family: 'JetBrains Mono', monospace; font-size: 0.95rem;
color: var(--accent4); margin: 1rem 0;
}
/* Comparison Table */
.comparison-table {
width: 100%; border-collapse: collapse; margin-top: 2rem;
background: var(--surface); border-radius: 14px; overflow: hidden;
border: 1px solid var(--border);
}
.comparison-table th, .comparison-table td {
padding: 1rem 1.5rem; text-align: left; font-size: 0.85rem;
border-bottom: 1px solid var(--border);
}
.comparison-table th {
background: var(--surface2); font-weight: 700; font-size: 0.75rem;
text-transform: uppercase; letter-spacing: 1px; color: var(--text-dim);
}
.comparison-table tr:last-child td { border-bottom: none; }
.comparison-table .highlight-row { background: rgba(108,99,255,0.08); }
.tag {
display: inline-block; padding: 0.2rem 0.6rem; border-radius: 6px;
font-size: 0.7rem; font-weight: 600;
}
.tag-green { background: rgba(67,233,123,0.15); color: var(--accent3); }
.tag-purple { background: rgba(108,99,255,0.15); color: var(--accent); }
.tag-pink { background: rgba(255,101,132,0.15); color: var(--accent2); }
/* Applications */
.app-grid {
display: grid; grid-template-columns: repeat(auto-fit, minmax(240px, 1fr));
gap: 1.5rem;
}
.app-card {
background: var(--surface); border: 1px solid var(--border);
border-radius: 16px; padding: 1.8rem; text-align: center;
transition: all 0.3s; cursor: default;
}
.app-card:hover { border-color: var(--accent2); transform: translateY(-3px); }
.app-icon { font-size: 2.5rem; margin-bottom: 1rem; }
.app-card h4 { font-size: 1rem; font-weight: 700; margin-bottom: 0.5rem; }
.app-card p { font-size: 0.8rem; color: var(--text-dim); }
/* Scroll animations */
.fade-up {
opacity: 0; transform: translateY(40px);
transition: opacity 0.7s ease, transform 0.7s ease;
}
.fade-up.visible { opacity: 1; transform: translateY(0); }
/* Divider */
.divider {
height: 1px; max-width: 1100px; margin: 0 auto;
background: linear-gradient(90deg, transparent, var(--border), transparent);
}
/* Footer */
footer {
border-top: 1px solid var(--border);
padding: 2rem;
}
.footer-inner {
max-width: 1200px; margin: 0 auto;
display: flex; justify-content: space-between; align-items: center;
flex-wrap: wrap; gap: 1rem;
}
.footer-inner span { font-size: 0.85rem; color: var(--text-dim); }
.footer-inner a { font-size: 0.85rem; color: var(--accent); text-decoration: none; transition: color 0.3s; }
.footer-inner a:hover { color: var(--text); }
/* Responsive */
@media (max-width: 768px) {
nav .nav-links { display: none; }
.hamburger { display: flex; }
nav .nav-links.open {
display: flex; flex-direction: column;
position: absolute; top: 64px; left: 0; right: 0;
background: rgba(10,10,15,0.98); padding: 2rem;
border-bottom: 1px solid var(--border);
}
.hero-stats { gap: 1.5rem; }
.demo-container { padding: 1.5rem; }
.comparison-table { font-size: 0.75rem; }
.comparison-table th, .comparison-table td { padding: 0.7rem; }
}
/* Glow pulse */
@keyframes glow-pulse {
0%, 100% { box-shadow: 0 0 20px rgba(108,99,255,0.2); }
50% { box-shadow: 0 0 40px rgba(108,99,255,0.4); }
}
.slider-container { display: flex; align-items: center; gap: 1rem; flex-wrap: wrap; }
.slider-container label { font-size: 0.85rem; font-weight: 600; }
input[type="range"] {
-webkit-appearance: none; appearance: none;
width: 180px; height: 6px; border-radius: 3px;
background: var(--border); outline: none;
}
input[type="range"]::-webkit-slider-thumb {
-webkit-appearance: none; appearance: none;
width: 18px; height: 18px; border-radius: 50%;
background: var(--accent); cursor: pointer;
}
.slider-val {
font-family: 'JetBrains Mono', monospace;
font-size: 0.85rem; color: var(--accent4);
min-width: 30px;
}
/* Phase indicator */
.phase-indicator {
display: flex; gap: 0; margin: 1.5rem 0; border-radius: 10px; overflow: hidden;
border: 1px solid var(--border);
}
.phase-seg {
padding: 0.7rem 1.2rem; font-size: 0.75rem; font-weight: 700;
text-transform: uppercase; letter-spacing: 1px; flex: 1;
text-align: center; transition: all 0.5s;
}
.phase-seg.explore-phase { background: rgba(108,99,255,0.15); color: var(--accent); }
.phase-seg.exploit-phase { background: rgba(67,233,123,0.08); color: var(--text-dim); }
.phase-seg.active-phase { background: rgba(108,99,255,0.3); }
.phase-seg.exploit-active { background: rgba(67,233,123,0.25); color: var(--accent3); }
.round-counter {
font-family: 'JetBrains Mono', monospace;
font-size: 1.5rem; font-weight: 700; color: var(--accent);
margin-bottom: 0.5rem;
}
.info-tooltip {
display: inline-flex; align-items: center; justify-content: center;
width: 18px; height: 18px; border-radius: 50%;
background: var(--border); color: var(--text-dim);
font-size: 0.65rem; font-weight: 700; cursor: help;
margin-left: 0.3rem; position: relative;
}
.info-tooltip:hover::after {
content: attr(data-tip);
position: absolute; bottom: 130%; left: 50%; transform: translateX(-50%);
padding: 0.5rem 0.8rem; border-radius: 8px; background: var(--surface2);
border: 1px solid var(--border); font-size: 0.7rem; color: var(--text);
white-space: nowrap; z-index: 10; font-weight: 400;
}
</style>
</head>
<body>
<!-- NAV -->
<nav id="navbar">
<div class="nav-inner">
<div class="logo">MA-CMAB</div>
<ul class="nav-links" id="nav-links">
<li><a href="#problem">The Problem</a></li>
<li><a href="#how">How It Works</a></li>
<li><a href="#demo">Live Demo</a></li>
<li><a href="#results">Results</a></li>
<li><a href="#applications">Applications</a></li>
</ul>
<button class="hamburger" id="hamburger" aria-label="Menu">
<span></span><span></span><span></span>
</button>
</div>
</nav>
<!-- HERO -->
<section class="hero">
<div class="hero-bg"></div>
<canvas id="hero-canvas"></canvas>
<div class="hero-content fade-up">
<div class="hero-badge">AAMAS 2026 · Accepted Paper</div>
<h1>Smarter Resource Sharing<br/>Through <span class="gradient-text">Learning Under Uncertainty</span></h1>
<p>How do you divide resources among multiple agents to maximize everyone's benefit—when you can only see the <em>total</em> outcome, not individual contributions? This paper solves it.</p>
<a href="#problem" class="hero-cta">Explore the Research ↓</a>
<div class="hero-stats">
<div class="hero-stat">
<div class="num">M</div>
<div class="label">Non-Communicating Agents</div>
</div>
<div class="hero-stat">
<div class="num">N</div>
<div class="label">Indivisible Items</div>
</div>
<div class="hero-stat">
<div class="num">T</div>
<div class="label">Learning Rounds</div>
</div>
<div class="hero-stat">
<div class="num">Õ(T<sup style="font-size:0.6em">⅔</sup>)</div>
<div class="label">Regret Bound</div>
</div>
</div>
</div>
</section>
<!-- THE PROBLEM -->
<div class="divider"></div>
<section id="problem">
<div class="section-inner fade-up">
<div class="section-label">The Problem</div>
<h2 class="section-title">Dividing Items Fairly When You Can't See Everything</h2>
<p class="section-desc">
Imagine you're an auctioneer dividing items among bidders, a platform assigning tasks to workers, or a system allocating bandwidth to users. Each recipient values bundles of items differently. Your goal: maximize <strong>total welfare</strong> — the sum of everyone's satisfaction.
</p>
<div class="card-grid">
<div class="card">
<div class="card-icon" style="background: rgba(108,99,255,0.15);">🧩</div>
<h3>Submodular Welfare Problem</h3>
<p>Each agent has a utility function with <strong>diminishing returns</strong> — getting more items helps, but each extra item helps a little less. This "submodularity" makes the math tractable but optimization hard.</p>
</div>
<div class="card">
<div class="card-icon" style="background: rgba(255,101,132,0.15);">🔒</div>
<h3>Bandit Feedback Only</h3>
<p>You only see the <strong>total reward</strong> after each allocation — not how much each agent individually benefited. It's like knowing a team's total score but not who scored each point.</p>
</div>
<div class="card">
<div class="card-icon" style="background: rgba(67,233,123,0.15);">🤐</div>
<h3>Non-Communicating Agents</h3>
<p>Agents <strong>can't talk to each other</strong>. There's no sharing of strategies, rewards, or information. Each operates independently, yet they compete for the same pool of items.</p>
</div>
<div class="card">
<div class="card-icon" style="background: rgba(56,249,215,0.15);">✂️</div>
<h3>Partition Constraint</h3>
<p>Each item goes to <strong>exactly one agent</strong> — no sharing or duplicating. This creates a complex partition matroid constraint that couples all agents' decisions together.</p>
</div>
</div>
</div>
</section>
<!-- HOW IT WORKS -->
<div class="divider"></div>
<section id="how">
<div class="section-inner fade-up">
<div class="section-label">The Solution</div>
<h2 class="section-title">Explore, Learn, Then Commit</h2>
<p class="section-desc">
The MA-CMAB algorithm uses a two-phase strategy: first explore different allocations to learn about utilities, then commit to the best one found. Simple in concept, powerful in practice.
</p>
<div style="display: grid; grid-template-columns: 1fr 1fr; gap: 3rem; align-items: start;" class="algo-layout">
<div>
<div class="algo-timeline" id="algo-timeline">
<div class="algo-step active" data-step="1">
<h4>1. Initialize</h4>
<p>Set up the exploration budget <em>m</em> based on the time horizon T, number of agents M, and items N. The key insight: explore for ~T<sup>2/3</sup> rounds.</p>
</div>
<div class="algo-step" data-step="2">
<h4>2. Exploration Phase</h4>
<p>Try different allocations. For each one, play it <em>m</em> times and average the rewards to get a reliable estimate of its true welfare value.</p>
</div>
<div class="algo-step" data-step="3">
<h4>3. Offline Oracle</h4>
<p>Feed the averaged estimates into the <strong>Continuous Greedy</strong> algorithm — a powerful offline method that finds a (1−1/e) ≈ 63% approximation of optimal welfare.</p>
</div>
<div class="algo-step" data-step="4">
<h4>4. Pipage Rounding</h4>
<p>Convert the fractional solution (e.g., "assign item 3 to agent 1 with probability 0.7") into a concrete integer allocation with almost no loss in quality.</p>
</div>
<div class="algo-step" data-step="5">
<h4>5. Exploitation Phase</h4>
<p>Commit to the best allocation found and play it for all remaining rounds. The exploration paid off — now reap the rewards.</p>
</div>
</div>
</div>
<div>
<div style="background: var(--surface); border: 1px solid var(--border); border-radius: 16px; padding: 2rem; position: sticky; top: 80px;">
<h4 style="margin-bottom: 1rem; font-size: 0.95rem;">Key Equation: Regret Bound</h4>
<div class="eq-box" style="display: block; text-align: center; font-size: 1.1rem;">
E[R(T)] = Õ(δ<sup>2/3</sup> · η<sup>1/3</sup> · C<sup>2/3</sup> · T<sup>2/3</sup>)
</div>
<div style="margin-top: 1.5rem;">
<div style="display: flex; align-items: start; gap: 0.8rem; margin-bottom: 0.8rem;">
<span class="tag tag-purple">α</span>
<span style="font-size: 0.85rem; color: var(--text-dim);">= 1 − 1/e ≈ 0.632 — the offline approximation ratio</span>
</div>
<div style="display: flex; align-items: start; gap: 0.8rem; margin-bottom: 0.8rem;">
<span class="tag tag-pink">δ</span>
<span style="font-size: 0.85rem; color: var(--text-dim);">= (4MN + 2M) — noise sensitivity parameter</span>
</div>
<div style="display: flex; align-items: start; gap: 0.8rem; margin-bottom: 0.8rem;">
<span class="tag tag-green">η</span>
<span style="font-size: 0.85rem; color: var(--text-dim);">= (MN)<sup>8</sup> — oracle complexity (# queries)</span>
</div>
</div>
<p style="font-size: 0.8rem; color: var(--text-dim); margin-top: 1rem; line-height: 1.6;">
The regret grows as T<sup>2/3</sup> — <strong>sublinearly</strong>. This means the average per-round regret → 0 as T → ∞. The algorithm eventually performs nearly as well as the best offline solution!
</p>
</div>
</div>
</div>
</div>
</section>
<!-- LIVE DEMO -->
<div class="divider"></div>
<section id="demo">
<div class="section-inner fade-up">
<div class="section-label">Interactive Demo</div>
<h2 class="section-title">Watch the Algorithm Learn in Real Time</h2>
<p class="section-desc">
See how the explore-then-commit strategy works. Adjust parameters and watch agents learn to allocate items for maximum welfare.
</p>
<div class="demo-container">
<div class="demo-header">
<div>
<h3>MA-CMAB Allocation Simulator</h3>
<p style="font-size: 0.8rem; color: var(--text-dim);">Agents have submodular utilities — each item helps, but with diminishing returns</p>
</div>
<div style="display: flex; gap: 0.5rem;">
<button class="btn btn-primary" id="btn-run">▶ Run Simulation</button>
<button class="btn btn-outline" id="btn-reset">↻ Reset</button>
</div>
</div>
<div class="slider-container" style="margin-bottom: 1.5rem; gap: 2rem;">
<div class="slider-container">
<label>Agents (M):</label>
<input type="range" id="slider-agents" min="2" max="5" value="3" />
<span class="slider-val" id="val-agents">3</span>
</div>
<div class="slider-container">
<label>Items (N):</label>
<input type="range" id="slider-items" min="4" max="12" value="6" />
<span class="slider-val" id="val-items">6</span>
</div>
<div class="slider-container">
<label>Rounds (T):</label>
<input type="range" id="slider-rounds" min="50" max="500" value="200" step="10" />
<span class="slider-val" id="val-rounds">200</span>
</div>
</div>
<div class="phase-indicator" id="phase-indicator">
<div class="phase-seg explore-phase active-phase" id="phase-explore">🔍 Exploration</div>
<div class="phase-seg exploit-phase" id="phase-exploit">🎯 Exploitation</div>
</div>
<div style="display: flex; align-items: center; gap: 1rem; margin-bottom: 1.5rem;">
<span style="font-size: 0.8rem; color: var(--text-dim);">Round:</span>
<span class="round-counter" id="round-counter">0 / 200</span>
</div>
<div id="alloc-viz"></div>
<div style="display: grid; grid-template-columns: 1fr 1fr; gap: 2rem;">
<div>
<div class="welfare-bar-label">Current Total Welfare <span class="info-tooltip" data-tip="Sum of all agents' submodular utilities">?</span></div>
<div class="welfare-bar-track"><div class="welfare-bar-fill" id="welfare-fill" style="width: 0%"></div></div>
<div class="welfare-value" id="welfare-value">0.00</div>
</div>
<div>
<div class="welfare-bar-label">Best Possible (1−1/e)·OPT <span class="info-tooltip" data-tip="The theoretical best the algorithm can guarantee">?</span></div>
<div class="welfare-bar-track"><div class="welfare-bar-fill" id="opt-fill" style="width: 63%; background: linear-gradient(90deg, var(--accent2), #ff9a9e);"></div></div>
<div class="welfare-value" id="opt-value" style="background: linear-gradient(135deg, var(--accent2), #ff9a9e); -webkit-background-clip: text;">≈ 0.63</div>
</div>
</div>
<div style="margin-top: 2rem;">
<div class="welfare-bar-label">Cumulative Regret Over Time</div>
<canvas id="regret-canvas"></canvas>
</div>
</div>
</div>
</section>
<!-- RESILIENCE -->
<div class="divider"></div>
<section id="results">
<div class="section-inner fade-up">
<div class="section-label">Key Results</div>
<h2 class="section-title">Robust Even With Noisy Information</h2>
<p class="section-desc">
A major contribution is proving <strong>resilience</strong>: the Continuous Greedy algorithm still works well even when utility estimates are noisy — which is exactly what happens during online learning.
</p>
<div class="card-grid" style="grid-template-columns: 1fr 1fr;">
<div class="card" style="border-color: rgba(108,99,255,0.3);">
<h3>Resilience Theorem</h3>
<p style="margin-bottom: 1rem;">Under noisy oracles with error ≤ ε, the Continuous Greedy algorithm achieves:</p>
<div class="eq-box" style="font-size: 0.85rem;">
f(S) ≥ (1 − 1/e) · f(OPT) − (4MN + 2M)·ε
</div>
<p style="margin-top: 1rem; font-size: 0.85rem; color: var(--text-dim);">
The approximation degrades <em>gracefully</em> — only linearly in noise ε. When ε → 0, we recover the classic (1−1/e) guarantee.
</p>
</div>
<div class="card" style="border-color: rgba(67,233,123,0.3);">
<h3>Clean Event Guarantee</h3>
<p style="margin-bottom: 1rem;">With high probability (≥ 1 − 2η/T), all empirical estimates concentrate around truth:</p>
<div class="eq-box" style="font-size: 0.85rem;">
|f̄(S) − f(S)| &lt; √(log T / 2m)
</div>
<p style="margin-top: 1rem; font-size: 0.85rem; color: var(--text-dim);">
This "clean event" makes the offline-to-online transfer rigorous. Bad estimates are exponentially unlikely.
</p>
</div>
</div>
<h3 style="margin-top: 3rem; margin-bottom: 1rem; font-size: 1.1rem;">How This Compares</h3>
<div style="overflow-x: auto;">
<table class="comparison-table">
<thead>
<tr>
<th>Framework</th>
<th>Agents</th>
<th>Communication</th>
<th>Action Type</th>
<th>Feedback</th>
<th>Regret</th>
</tr>
</thead>
<tbody>
<tr>
<td>C-ETC (Nie et al., 2023)</td>
<td>Single</td>
<td>N/A</td>
<td>Single subset</td>
<td>Full-bandit</td>
<td>Õ(T<sup>2/3</sup>)</td>
</tr>
<tr>
<td>Fourati et al. (2024a)</td>
<td>Multi</td>
<td><span class="tag tag-pink">Yes</span></td>
<td>Shared subset</td>
<td>Semi-bandit</td>
<td>Õ(T<sup>2/3</sup>)</td>
</tr>
<tr class="highlight-row">
<td><strong>MA-CMAB (This Paper)</strong></td>
<td><strong>Multi</strong></td>
<td><span class="tag tag-green">None</span></td>
<td><strong>Partition</strong></td>
<td><strong>Full-bandit</strong></td>
<td><strong>Õ(T<sup>2/3</sup>)</strong></td>
</tr>
</tbody>
</table>
</div>
<p style="font-size: 0.8rem; color: var(--text-dim); margin-top: 0.8rem;">
★ MA-CMAB achieves the same regret rate with the hardest feedback model (full-bandit) and no inter-agent communication — a strictly harder setting.
</p>
</div>
</section>
<!-- APPLICATIONS -->
<div class="divider"></div>
<section id="applications">
<div class="section-inner fade-up">
<div class="section-label">Real-World Impact</div>
<h2 class="section-title">Where This Framework Applies</h2>
<p class="section-desc">
The MA-CMAB framework applies wherever you need to divide resources among competing parties and learn from aggregate outcomes.
</p>
<div class="app-grid">
<div class="app-card">
<div class="app-icon">🛒</div>
<h4>Combinatorial Auctions</h4>
<p>Allocate goods to bidders in repeated auctions where true valuations are unknown and only total revenue is observed.</p>
</div>
<div class="app-card">
<div class="app-icon">📱</div>
<h4>Recommender Systems</h4>
<p>Partition content across user groups to maximize total engagement, with only aggregate click-through data available.</p>
</div>
<div class="app-card">
<div class="app-icon">📡</div>
<h4>Distributed Sensing</h4>
<p>Assign sensor targets to non-communicating agents to maximize total coverage with only aggregate quality feedback.</p>
</div>
<div class="app-card">
<div class="app-icon">🌐</div>
<h4>Influence Maximization</h4>
<p>Divide seed users across campaigns to maximize total influence spread in social networks.</p>
</div>
<div class="app-card">
<div class="app-icon">⚖️</div>
<h4>Fair Division</h4>
<p>Allocate indivisible goods among people with unknown preferences to maximize utilitarian social welfare.</p>
</div>
<div class="app-card">
<div class="app-icon">📊</div>
<h4>Data Summarization</h4>
<p>Partition datasets across workers to maximize total information coverage with submodular diversity objectives.</p>
</div>
</div>
</div>
</section>
<!-- TAKEAWAY -->
<div class="divider"></div>
<section>
<div class="section-inner fade-up" style="text-align: center;">
<div class="section-label">Key Takeaway</div>
<h2 class="section-title" style="max-width: 800px; margin: 0 auto 1.5rem;">
The First Sublinear Regret Guarantee for<br/>
<span style="background: linear-gradient(135deg, var(--accent), var(--accent3)); -webkit-background-clip: text; -webkit-text-fill-color: transparent;">
Multi-Agent Partitioned Welfare Maximization
</span><br/>Under Bandit Feedback
</h2>
<p style="color: var(--text-dim); max-width: 600px; margin: 0 auto 2rem; line-height: 1.7;">
By bridging offline submodular optimization with online multi-agent bandit learning, this work opens the door to practical, scalable algorithms for resource allocation under real-world information constraints.
</p>
<div class="eq-box" style="font-size: 1.1rem; animation: glow-pulse 3s ease-in-out infinite;">
Regret → 0 per round as T → ∞
</div>
</div>
</section>
<!-- FOOTER -->
<footer>
<div class="footer-inner">
<span>© 2026 Amroja, LLC</span>
<a href="https://johndamask.com" target="_blank" rel="noopener">johndamask.com</a>
</div>
</footer>
<script>
// --- NAV ---
const hamburger = document.getElementById('hamburger');
const navLinks = document.getElementById('nav-links');
hamburger.addEventListener('click', () => navLinks.classList.toggle('open'));
document.querySelectorAll('.nav-links a').forEach(a => a.addEventListener('click', () => navLinks.classList.remove('open')));
// --- SCROLL ANIMATIONS ---
const fadeEls = document.querySelectorAll('.fade-up');
const obs = new IntersectionObserver((entries) => {
entries.forEach(e => { if (e.isIntersecting) { e.target.classList.add('visible'); obs.unobserve(e.target); } });
}, { threshold: 0.15 });
fadeEls.forEach(el => obs.observe(el));
// --- ALGO TIMELINE ---
const steps = document.querySelectorAll('.algo-step');
let currentStep = 0;
setInterval(() => {
steps.forEach(s => s.classList.remove('active'));
steps[currentStep].classList.add('active');
currentStep = (currentStep + 1) % steps.length;
}, 2500);
// --- HERO PARTICLES ---
(function() {
const canvas = document.getElementById('hero-canvas');
const ctx = canvas.getContext('2d');
let particles = [];
function resize() { canvas.width = canvas.offsetWidth; canvas.height = canvas.offsetHeight; }
resize();
window.addEventListener('resize', resize);
for (let i = 0; i < 60; i++) {
particles.push({
x: Math.random() * canvas.width,
y: Math.random() * canvas.height,
vx: (Math.random() - 0.5) * 0.3,
vy: (Math.random() - 0.5) * 0.3,
r: Math.random() * 2 + 1,
color: ['rgba(108,99,255,','rgba(255,101,132,','rgba(67,233,123,','rgba(56,249,215,'][Math.floor(Math.random()*4)]
});
}
function draw() {
ctx.clearRect(0, 0, canvas.width, canvas.height);
particles.forEach((p, i) => {
p.x += p.vx; p.y += p.vy;
if (p.x < 0 || p.x > canvas.width) p.vx *= -1;
if (p.y < 0 || p.y > canvas.height) p.vy *= -1;
ctx.beginPath();
ctx.arc(p.x, p.y, p.r, 0, Math.PI * 2);
ctx.fillStyle = p.color + '0.5)';
ctx.fill();
// Connect nearby particles
for (let j = i + 1; j < particles.length; j++) {
const dx = p.x - particles[j].x, dy = p.y - particles[j].y;
const dist = Math.sqrt(dx*dx + dy*dy);
if (dist < 120) {
ctx.beginPath();
ctx.moveTo(p.x, p.y);
ctx.lineTo(particles[j].x, particles[j].y);
ctx.strokeStyle = p.color + (0.15 * (1 - dist/120)) + ')';
ctx.lineWidth = 0.5;
ctx.stroke();
}
}
});
requestAnimationFrame(draw);
}
draw();
})();
// --- SIMULATION ---
const agentColors = ['#6c63ff', '#ff6584', '#43e97b', '#38f9d7', '#ffd700'];
const agentNames = ['Alice', 'Bob', 'Carol', 'Dave', 'Eve'];
const itemEmojis = ['📦','💎','🎨','📚','🔧','🎵','🏀','🧪','🌸','🎯','⚡','🍕'];
let simState = null;
let simRunning = false;
let simInterval = null;
function getParams() {
return {
M: +document.getElementById('slider-agents').value,
N: +document.getElementById('slider-items').value,
T: +document.getElementById('slider-rounds').value
};
}
// Submodular utility: sqrt-based coverage
function submodularUtil(items, agentIdx, N) {
if (items.length === 0) return 0;
// Each agent values items differently
let val = 0;
const seen = new Set();
items.forEach(item => {
if (!seen.has(item)) {
seen.add(item);
// Weighted diminishing returns
const weight = ((item + agentIdx * 3 + 1) % N + 1) / N;
val += weight / Math.sqrt(seen.size);
}
});
return Math.min(val / Math.sqrt(N), 1);
}
function totalWelfare(allocation, M, N) {
let total = 0;
for (let i = 0; i < M; i++) {
total += submodularUtil(allocation[i], i, N);
}
return total;
}
function randomPartition(M, N) {
const alloc = Array.from({length: M}, () => []);
const items = Array.from({length: N}, (_, i) => i);
// Shuffle
for (let i = items.length - 1; i > 0; i--) {
const j = Math.floor(Math.random() * (i + 1));
[items[i], items[j]] = [items[j], items[i]];
}
items.forEach((item, idx) => alloc[idx % M].push(item));
return alloc;
}
function greedyPartition(M, N) {
const alloc = Array.from({length: M}, () => []);
const items = Array.from({length: N}, (_, i) => i);
// Greedy: for each item, assign to agent with max marginal gain
items.forEach(item => {
let bestAgent = 0, bestGain = -1;
for (let i = 0; i < M; i++) {
const gain = submodularUtil([...alloc[i], item], i, N) - submodularUtil(alloc[i], i, N);
if (gain > bestGain) { bestGain = gain; bestAgent = i; }
}
alloc[bestAgent].push(item);
});
return alloc;
}
function initSim() {
const {M, N, T} = getParams();
const optAlloc = greedyPartition(M, N);
const optWelfare = totalWelfare(optAlloc, M, N);
const exploreRounds = Math.ceil(Math.pow(T, 2/3));
simState = {
M, N, T, round: 0,
exploreRounds,
phase: 'explore',
bestAlloc: null,
bestWelfare: 0,
currentAlloc: Array.from({length: M}, () => []),
optWelfare,
regretHistory: [],
cumRegret: 0,
triedAllocs: [],
triedWelfares: [],
currentWelfare: 0
};
renderAllocViz();
updateDisplay();
drawRegretChart();
}
function renderAllocViz() {
const {M, N} = getParams();
const viz = document.getElementById('alloc-viz');
viz.innerHTML = '';
for (let i = 0; i < M; i++) {
const box = document.createElement('div');
box.className = 'agent-box';
box.id = `agent-box-${i}`;
box.innerHTML = `
<div class="agent-label">
<span class="agent-dot" style="background:${agentColors[i]}"></span>
${agentNames[i]} (Agent ${i+1})
</div>
<div class="agent-items" id="agent-items-${i}"></div>
<div style="margin-top:0.8rem;">
<span style="font-size:0.7rem;color:var(--text-dim)">Utility:</span>
<span style="font-family:'JetBrains Mono';font-size:0.85rem;color:${agentColors[i]}" id="agent-util-${i}">0.00</span>
</div>
`;
viz.appendChild(box);
}
}
function updateAllocViz(alloc) {
const {M, N} = simState;
for (let i = 0; i < M; i++) {
const container = document.getElementById(`agent-items-${i}`);
const box = document.getElementById(`agent-box-${i}`);
if (!container) continue;
container.innerHTML = '';
(alloc[i] || []).forEach(item => {
const chip = document.createElement('span');
chip.className = 'item-chip';
chip.textContent = `${itemEmojis[item % itemEmojis.length]} Item ${item+1}`;
container.appendChild(chip);
});
const util = submodularUtil(alloc[i] || [], i, N);
const utilEl = document.getElementById(`agent-util-${i}`);
if (utilEl) utilEl.textContent = util.toFixed(3);
box.classList.remove('highlight');
if (alloc[i] && alloc[i].length > 0) {
box.classList.add('highlight');
setTimeout(() => box.classList.remove('highlight'), 600);
}
}
}
function updateDisplay() {
const s = simState;
if (!s) return;
document.getElementById('round-counter').textContent = `${s.round} / ${s.T}`;
const expPhase = document.getElementById('phase-explore');
const expltPhase = document.getElementById('phase-exploit');
expPhase.className = 'phase-seg explore-phase' + (s.phase === 'explore' ? ' active-phase' : '');
expltPhase.className = 'phase-seg exploit-phase' + (s.phase === 'exploit' ? ' exploit-active' : '');
document.getElementById('welfare-fill').style.width = (s.currentWelfare / Math.max(s.optWelfare, 0.01) * 100) + '%';
document.getElementById('welfare-value').textContent = s.currentWelfare.toFixed(3);
const approxOpt = (1 - 1/Math.E) * s.optWelfare;
document.getElementById('opt-fill').style.width = (approxOpt / Math.max(s.optWelfare, 0.01) * 100) + '%';
document.getElementById('opt-value').textContent = `≈ ${approxOpt.toFixed(3)}`;
}
function simStep() {
const s = simState;
if (!s || s.round >= s.T) {
stopSim();
return;
}
s.round++;
const {M, N} = s;
const benchmark = (1 - 1/Math.E) * s.optWelfare;
if (s.round <= s.exploreRounds) {
s.phase = 'explore';
const alloc = randomPartition(M, N);
const w = totalWelfare(alloc, M, N) + (Math.random() - 0.5) * 0.05; // noise
s.triedAllocs.push(alloc);
s.triedWelfares.push(w);
if (w > s.bestWelfare) {
s.bestWelfare = w;
s.bestAlloc = alloc;
}
s.currentAlloc = alloc;
s.currentWelfare = Math.max(0, w);
} else {
s.phase = 'exploit';
if (!s.committed) {
// Run greedy on best estimates
s.bestAlloc = greedyPartition(M, N);
s.bestWelfare = totalWelfare(s.bestAlloc, M, N);
s.committed = true;
}
const noise = (Math.random() - 0.5) * 0.02;
s.currentAlloc = s.bestAlloc;
s.currentWelfare = Math.max(0, s.bestWelfare + noise);
}
const instantRegret = Math.max(0, benchmark - s.currentWelfare);
s.cumRegret += instantRegret;
s.regretHistory.push(s.cumRegret);
updateAllocViz(s.currentAlloc);
updateDisplay();
drawRegretChart();
}
function startSim() {
if (simRunning) return;
initSim();
simRunning = true;
const speed = simState.T > 200 ? 20 : 40;
simInterval = setInterval(simStep, speed);
}
function stopSim() {
simRunning = false;
if (simInterval) clearInterval(simInterval);
simInterval = null;
}
function resetSim() {
stopSim();
initSim();
}
document.getElementById('btn-run').addEventListener('click', startSim);
document.getElementById('btn-reset').addEventListener('click', resetSim);
['slider-agents','slider-items','slider-rounds'].forEach(id => {
const el = document.getElementById(id);
const valId = id.replace('slider-', 'val-');
el.addEventListener('input', () => {
document.getElementById(valId).textContent = el.value;
if (!simRunning) initSim();
});
});
// --- REGRET CHART ---
function drawRegretChart() {
const canvas = document.getElementById('regret-canvas');
const ctx = canvas.getContext('2d');
const dpr = window.devicePixelRatio || 1;
const rect = canvas.getBoundingClientRect();
canvas.width = rect.width * dpr;
canvas.height = rect.height * dpr;
ctx.scale(dpr, dpr);
const W = rect.width, H = rect.height;
ctx.clearRect(0, 0, W, H);
if (!simState || simState.regretHistory.length < 2) {
ctx.fillStyle = '#8888a0';
ctx.font = '14px Inter';
ctx.textAlign = 'center';
ctx.fillText('Run the simulation to see regret curve', W/2, H/2);
return;
}
const data = simState.regretHistory;
const T = simState.T;
const maxR = Math.max(...data, 1);
const pad = {t: 30, r: 20, b: 40, l: 60};
const cW = W - pad.l - pad.r;
const cH = H - pad.t - pad.b;
// Grid lines
ctx.strokeStyle = 'rgba(42,42,62,0.6)';
ctx.lineWidth = 0.5;
for (let i = 0; i <= 4; i++) {
const y = pad.t + (cH / 4) * i;
ctx.beginPath(); ctx.moveTo(pad.l, y); ctx.lineTo(W - pad.r, y); ctx.stroke();
ctx.fillStyle = '#8888a0'; ctx.font = '10px JetBrains Mono'; ctx.textAlign = 'right';
ctx.fillText((maxR * (1 - i/4)).toFixed(1), pad.l - 8, y + 4);
}
// T^{2/3} reference curve
ctx.strokeStyle = 'rgba(255,101,132,0.3)';
ctx.lineWidth = 1.5;
ctx.setLineDash([4, 4]);
ctx.beginPath();
const refScale = maxR / Math.pow(data.length, 2/3);
for (let i = 0; i < data.length; i++) {
const x = pad.l + (i / T) * cW;
const y = pad.t + cH - (refScale * Math.pow(i+1, 2/3) / maxR) * cH;
if (i === 0) ctx.moveTo(x, y); else ctx.lineTo(x, y);
}
ctx.stroke();
ctx.setLineDash([]);
// Exploration region
if (simState.exploreRounds > 0) {
const exX = pad.l + (simState.exploreRounds / T) * cW;
ctx.fillStyle = 'rgba(108,99,255,0.06)';
ctx.fillRect(pad.l, pad.t, exX - pad.l, cH);
ctx.strokeStyle = 'rgba(108,99,255,0.3)';
ctx.lineWidth = 1;
ctx.setLineDash([3,3]);
ctx.beginPath(); ctx.moveTo(exX, pad.t); ctx.lineTo(exX, pad.t + cH); ctx.stroke();
ctx.setLineDash([]);
ctx.fillStyle = 'rgba(108,99,255,0.6)'; ctx.font = '10px Inter'; ctx.textAlign = 'center';
ctx.fillText('Explore', (pad.l + exX)/2, pad.t + 15);
ctx.fillStyle = 'rgba(67,233,123,0.5)';
ctx.fillText('Exploit', (exX + W - pad.r)/2, pad.t + 15);
}
// Main curve
const grad = ctx.createLinearGradient(pad.l, 0, W - pad.r, 0);
grad.addColorStop(0, '#6c63ff');
grad.addColorStop(0.5, '#43e97b');
grad.addColorStop(1, '#38f9d7');
ctx.strokeStyle = grad;
ctx.lineWidth = 2.5;
ctx.beginPath();
data.forEach((v, i) => {
const x = pad.l + (i / T) * cW;
const y = pad.t + cH - (v / maxR) * cH;
if (i === 0) ctx.moveTo(x, y); else ctx.lineTo(x, y);
});
ctx.stroke();
// Fill under curve
ctx.lineTo(pad.l + ((data.length - 1) / T) * cW, pad.t + cH);
ctx.lineTo(pad.l, pad.t + cH);
ctx.closePath();
const fillGrad = ctx.createLinearGradient(0, pad.t, 0, pad.t + cH);
fillGrad.addColorStop(0, 'rgba(108,99,255,0.15)');
fillGrad.addColorStop(1, 'rgba(108,99,255,0)');
ctx.fillStyle = fillGrad;
ctx.fill();
// Labels
ctx.fillStyle = '#8888a0'; ctx.font = '11px Inter'; ctx.textAlign = 'center';
ctx.fillText('Round', W/2, H - 8);
ctx.save();
ctx.translate(14, H/2);
ctx.rotate(-Math.PI/2);
ctx.fillText('Cumulative Regret', 0, 0);
ctx.restore();
// Legend
ctx.fillStyle = '#38f9d7'; ctx.font = '10px Inter'; ctx.textAlign = 'left';
ctx.fillRect(W - pad.r - 140, pad.t + 5, 10, 3);
ctx.fillText('Actual regret', W - pad.r - 125, pad.t + 10);
ctx.fillStyle = 'rgba(255,101,132,0.5)';
ctx.setLineDash([4,4]);
ctx.beginPath(); ctx.moveTo(W - pad.r - 140, pad.t + 25); ctx.lineTo(W - pad.r - 130, pad.t + 25); ctx.stroke();
ctx.setLineDash([]);
ctx.fillStyle = 'rgba(255,101,132,0.7)';
ctx.fillText('T^(2/3) reference', W - pad.r - 125, pad.t + 28);
}
// Init
initSim();
window.addEventListener('resize', () => { if (simState) drawRegretChart(); });
// --- RESPONSIVE ALGO LAYOUT ---
const algoLayout = document.querySelector('.algo-layout');
function checkAlgoLayout() {
if (window.innerWidth < 768) {
algoLayout.style.gridTemplateColumns = '1fr';
} else {
algoLayout.style.gridTemplateColumns = '1fr 1fr';
}
}
checkAlgoLayout();
window.addEventListener('resize', checkAlgoLayout);
</script>
</body>
</html>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment