colletti.r_
riccardo@berkeley:~/portfolio 
$ whoami
Hi, I'm Riccardo. Yes, another engineer, but
(mostly) fun at parties.

$ cat ./about.txt
I studied in Milan and now I'm in Berkeley.
Translation: I've moved around a lot, collected
too many student IDs, and somehow still haven't
figured out how to cook pasta properly (yes,
I know, shame on me, I'm Italian).

$ ls ./now/
hydramarl/   klip_ppo/   drift_happens/

$ echo $LOOKING_FOR
AI / ML / RL · PM · Growth · Product Marketing
→ places where data and gut both get to talk.

$ echo $INTERESTS
🎾 padel  ·  🏃 running  ·  ⚽ Inter Milan  ·  ✈️ travel

$ _
// 01

about

/* who, what, why */

Hi, I'm Riccardo. I do research in reinforcement learning, from the theory behind policy optimization to systems where many agents learn together. I've just finished my MEng in EECS at UC Berkeley, after a Master's at Bocconi and an engineering degree at Politecnico di Torino. Three countries, three degrees, and probably too much espresso.

Sports taught me discipline and how to take feedback without taking it personally. Engineering taught me rigor. Business taught me that rigor, on its own, is rarely enough. Volunteering reminded me that not everything is about moving faster, and that sometimes stopping to pay attention is part of doing meaningful work.

When something doesn't make sense or refuses to work, I tend to obsess over it until it gives in. Everything else, less so. I work best with people I can learn from, who raise the bar, and who make the work fun.

// riccardo.stack
fuel espresso · pasta · gelato
training padel (sponsored, top 200 italy) · running · swimming
team Inter Milan: for the highs, the lows, and the swearing in between
moves turin → milan → ireland → milan → berkeley
langs italian · english (ielts 8/9) · latin (rusty) · french (nodding politely)
volunteering OFTAL (transport coordinator) · Banco Alimentare · AGAPE Onlus
~/life/ 5 frames
01padel.jpg
02running.jpg
03ski_italy.jpg
04warriors.jpg
05party.jpg
~/life/travel/ 6 frames
01cambodia.jpg
02finland.jpg
03toronto.jpg
04golden_gate.jpg
05hollywood.jpg
06grand_canyon.jpg
// 02

current research

/* what I'm building right now */
~/research/hydramarl.md 2026 · neurips '26 main track · under review
a1a2 a3a4 shared trunk head 1head 2 head 3head 4 one parameter λ moves between the two regimes
actor & critic · shared trunk ⇄ per-agent heads λ = 1.00 · shared

HydraMARL

A foundational study of parameter sharing in cooperative multi-agent reinforcement learning. HydraMARL decomposes each agent's policy and value networks into a population-shared trunk and per-agent heads, applied symmetrically across actor and critic; a single continuous parameter spans the full spectrum from complete sharing to full independence. Systematic experiments on the VMAS benchmark suite identify the optimal sharing regime and characterise its sensitivity to task structure, degree of cooperation, and population size.

When many AI agents learn together, parts of their neural networks can be shared. We isolate how much sharing actually helps, and expose the real trade-off between learning fast and learning to specialize.

  • multi-agent rl
  • parameter sharing
  • neurips '26
~/research/klip_ppo.md 2026 · arXiv preprint · klip-ppo.org ↗
0 + − I_kill I_kill I_pass I_pass I_in 1−ε 1 1+ε importance ratio w → advantage Â
w–Â plane · where clipping suppresses the update I_in · β = 0

KLip-PPO

A theoretical unification of PPO-Clip and PPO-KL, the two dominant variants of the most widely used policy-gradient algorithm in modern reinforcement learning. We prove a per-sample equivalence: the gradient of the clipped surrogate is exactly a KL penalty whose coefficient varies per sample, in closed form in the importance ratio and the advantage. The identity holds at every update step, and on five MuJoCo benchmarks the two losses trace indistinguishable training curves. The result collapses what was treated as a divide in the field into a single design space, and uncovers a principled family of generalizations of PPO.

PPO is the most widely used algorithm in modern RL. Its two main variants were treated as different ideas: we prove they compute the same update, and verify the identity across five MuJoCo benchmarks.

  • rl theory
  • ppo
  • policy optimization
klip-ppo.org ↗ arXiv ↗ code ↗ w&b artifacts ↗
~/research/drift_happens.md 2026 · published @ qcds, ecml-pkdd '26 · drift-happens.org ↗
The mean yearbook portrait per decade, 1910s to 2010s.
~/yearbook/mean_face · mean portrait per decade 1910s
eval 1960 eval 1980 eval 2000 cnn_l
CNN saliency, eval 1960: attribution concentrates on facial features.
CNN saliency, eval 1980.
CNN saliency, eval 2000.
resnet_s
ResNet saliency, eval 1960.
ResNet saliency, eval 1980.
ResNet saliency, eval 2000.
mlp_l
MLP saliency, eval 1960: attribution spread over the whole image.
MLP saliency, eval 1980.
MLP saliency, eval 2000.
saliency · trained ≤ 1950 · where each net looks eval 1960

Drift Happens

A large-scale empirical study of how neural architectures hold up under temporal distribution shift. Using temporal drift matrices (models trained on cumulative history and evaluated on every period) we compare MLPs, CNNs, Transformers, and frozen pretrained encoders across a century of yearbook portraits, a decade of Amazon reviews, and 25 years of arXiv abstracts. In-distribution accuracy is not a reliable guide to temporal robustness: the architectures that dominate the training period often degrade fastest, while frozen pretrained encoders trade accuracy for stability.

AI models lose accuracy as the world drifts away from their training data. We measured that decay across decades of faces, reviews, and papers: the models that start strongest are often the first to degrade.

  • neural nets
  • distribution shift
  • empirical
drift-happens.org ↗ paper ↗ code ↗ w&b artifacts ↗
~/robotics/hold_your_plate.md 2025 · robotics · holdurpla7e.org ↗
~/sim/arm_hero.webm · policy rollout t=0.0s
Bar charts comparing PID and Residual PPO across 10 domain conditions: ball drop rate and tracking error.
pid vs residual ppo · 10 conditions · click to zoom drop 70→3.3%

Hold Your Plate

Real-time ball-on-plate stabilization on a UR7e collaborative arm, with ArUco vision tracking through an Intel RealSense camera, orchestrated on a full ROS 2 stack and a MuJoCo simulation pipeline. The system compares three controllers: a hand-tuned PID, an MPC on the linearized dynamics, and a residual learning approach that augments the PID with a small PPO policy trained under domain randomization (varied mass, friction, sensing delay), pairing the reliability of model-based control with learned robustness to unmodeled effects.

A UR7e robotic arm balancing a ball on a tilting plate, with a camera tracking it in real time. Compares three controllers (classical, model-based, and learned) on a ROS 2 stack.

  • ros 2
  • residual rl
  • ur7e
  • mujoco
  • vision
holdurpla7e.org ↗ code ↗
~/research/demand.md 2026 · framework

DEMAND / no-steady-state

A research framework for multi-agent reinforcement learning under non-stationary demand, targeting fleet dispatch problems in delivery, ride-hailing, and logistics. The platform models the full operational stack (order matching, routing, courier repositioning) under realistic exogenous drift, with plug-and-play integration of 15+ RL algorithms (PPO, SAC, multi-agent variants, classical heuristics). Designed for the controlled drift studies that current MARL benchmarks cannot support.

A research platform for studying how AI dispatch systems (food delivery, ride-hailing) hold up when demand keeps changing. Full simulation pipeline with many algorithms wired in.

  • multi-agent rl
  • simulation
  • operations
~/capstone/no_free_lunch.md 2026 · meng capstone thesis · informs/meituan

No Free Lunch for Food Delivery

A multi-objective study of courier-assignment policies on real Meituan delivery data, where customer latency, courier earnings, platform margin, and restaurant reliability fundamentally conflict. We benchmark classical heuristics and modern reinforcement learning, then propose PRISM: a hybrid architecture combining learned cost prediction, RL, and constraint optimization that treats Pareto trade-offs as first-class objectives instead of collapsing them into a scalar reward.

In food delivery, customers, couriers, platforms, and restaurants all want different things from the same decision. Built on real Meituan data, PRISM treats those trade-offs as first-class instead of pretending they aren't there.

  • operations research
  • multi-objective
  • real data
// 03

snapshots

/* a portrait of my life · the [EE] ones hide easter eggs */
~/life/quadri/ 5 frames · easter eggs inside
01quadro_01.jpgEE
02quadro_02.jpgEE
03quadro_03.jpgEE
04quadro_04.jpgEE
05quadro_05.jpgEE
// 04

cv

/* timeline */

# education

  1. Aug 2025 → May 2026
    UC Berkeley · Berkeley, CA GPA 4.0 / 4.0
    MEng in Electrical Engineering & Computer Sciences Statistical learning theory (CS C281A, Recht), machine learning (CS 289A), deep reinforcement learning (CS 285, Levine), optimization (EECS 227AT, Sojoudi), robotics (EECS C206A). Capstone advised by Prof. Paul Grigas (IEOR). Fung Excellence Scholar: Berkeley Engineering merit award for top MEng students.
  2. Sep 2023 → May 2025
    Bocconi University · Milan, IT 110L / 110 (cum laude)
    Master in Marketing & Communication (MiMeC) Valedictorian: best overall student of the cohort. Merit-based scholarship. Quantitative program focused on analytics, consumer behavior, pricing strategy, and data-informed decision-making.
  3. Sep 2017 → Jul 2021
    Politecnico di Torino · Turin, IT 110L / 110 (cum laude)
    BSc in Electronic & Communications Engineering Top of class · English-taught ICT engineering program. Honors track, Intraprendenti (Young Talents Program): top 200 students university-wide. Thesis on Digital PLL simulation.
  4. Aug 2015 → Jun 2016
    Coláiste Chríost Rí · Cork, IE
    Exchange year · English-speaking secondary school Full-immersion academic year alongside Irish peers, within the Liceo Scientifico Internazionale track. Played in the school basketball team.

# experience

  1. Jan → Jul 2025
    Lenovo · Milan, IT
    Marketing Specialist · CRM Digital Services CRM and marketing analytics: customer segmentation, campaign analysis, CRM data exploration. Translated analytical insights into marketing and growth inputs inside a global, data-intensive corporate environment.
  2. 2021 → 2025
    Rialco S.R.L. · Milan, IT
    CEO / Product Manager / Growth Manager Owned strategic direction and P&L of a digital consulting firm. Led discovery → ship for digital products (web, e-commerce, internal tooling). Ran acquisition and conversion across SEO, paid, and social, defining KPIs and using analytics to prioritize what moved the funnel.
  3. Jan → Jul 2020
    Biesse Sistemi S.R.L. · Nizza Monferrato, IT
    Intern · Software Development Voice-controlled application for automated taxi booking: application logic and system integration alongside a senior software developer.

# publications & preprints

  1. May 2026
    HydraMARL: Isolating the Parameter-Sharing Spectrum in Multi-Agent Reinforcement Learning under review
    Holzinger, R.* & Colletti, R.* Submitted to NeurIPS 2026, main track.
  2. Jun 2026
    KLip-PPO: A per-sample KL perspective on PPO-Clip preprint
    Colletti, R.* & Holzinger, R.* arXiv:2606.23932 · klip-ppo.org · code · w&b artifacts · NeurIPS 2026 workshop submission planned.
  3. Jun 2026
    Drift Happens: An Empirical Study of Neural Architecture Robustness to Temporal Distribution Shift published
    Holzinger, R.* & Colletti, R.* Published at QCDS @ ECML-PKDD 2026 · drift-happens.org · code · w&b artifacts.

* equal contribution

# awards & scholarships

  1. 2025
    Fung Excellence Scholarship · UC Berkeley Awarded by Berkeley Engineering to top MEng candidates.
  2. 2025
    Valedictorian + Merit-Based Scholarship · Bocconi University Best overall student of the cohort.
  3. 2025
    Data Science and Machine Learning · MIT IDSS Making Data-Driven Decisions · professional certificate program (May–Oct 2025).
  4. 2017 → 2021
    Intraprendenti (Young Talents Program) · Politecnico di Torino Top 200 students university-wide.
  5. 2016
    Italian Mathematical Olympiad, national finalist · team competition Flown back from the exchange year in Ireland for one week to compete at the national finals.

# volunteering

  1. 2010 → now
    OFTAL · volunteer, now Transport Coordinator 15+ years supporting pilgrimages that bring ill and economically disadvantaged people to Lourdes and other shrines.
  2. 2014 → 2018
    Fondazione Banco Alimentare Onlus Food selection and distribution to families in need.
  3. 2014
    AGAPE Onlus Food sorting and distribution for the local community.

$ open cv.pdf ↗ · resume.pdf ↗