Pedro Botelho

Research Assistant @ Northwestern & Carnegie Mellon
Research Scientist @ CEIA (UFG) 🇧🇷
(Multi-Agent) Reinforcement Learning, Robotics, and ML

[Pedro Botelho @PedroBotelhoRSc]

I'm happy to share that our paper has been accepted to Interspeech 2026!

CAL-MOS: Bridging Layers with Adapters for Robust MOS Prediction Across Speech Foundation Models

https://t.co/zY5LhTsI72

#Interspeech #Interspeech2026 #SSL #SSL #ML #AI #DeepLearning #Speech

[Pedro Botelho @PedroBotelhoRSc]

@solrewards 2 x 1 Mexico

[Pedro Botelho @PedroBotelhoRSc]

@MarcusCordeiro9 @oraulsena O governo de Goiás expandiu o CEIA (antiga iniciativa interna da UFG) para uma iniciativa estadual e há alguns anos o governo federal expandiu os investimentos para atender o Brasil.

[Pedro Botelho @PedroBotelhoRSc]

@oraulsena A maioria da galera do CEIA é da UFG, mas envolve inúmeros pesquisadores formados em outras universidades. Me formei no ITA e agora pesquiso com duas universidades americanas (Northwestern e Carnegie Mellon), mas também atuo pesquisando com o CEIA :)

[Pedro Botelho @PedroBotelhoRSc]

@oraulsena Para quem está curioso e não entendeu como funciona:
1 - Empresa contrata o CEIA para desenvolver um algoritmo/IA (Ifood, Globo e CEMIG são exemplos reais). Aportam parte da grana.
2 - Como o projeto é de risco governo federal e estudal aportam outra parte.
Vocês usam e não sabem

[Pedro Botelho @PedroBotelhoRSc]

@zucraldi @Castro1021 You're right 😅
I wrote it in Portuguese.

[Pedro Botelho @PedroBotelhoRSc]

I’m happy to share that I’ve been admitted to the Machine Learning Summer School NYC 2026, hosted at @Columbia. 😁 🎊
Looking forward to learning from leading researchers and deepening my understanding of topics such as Causal ML and RL.

#ML #MLSS #Columbia #AI

[Elias Bareinboim @eliasbareinboim]

Interesting thread. One key point that often gets overlooked: even if models learned continuously post-deployment, interaction alone doesn't resolve the core issue.

The same transition data can arise from different causal mechanisms that behave identically under observed policies but diverge under new interventions.

So the bottleneck isn't just "learning by doing," but identifying which mechanisms generate the data.

This came up in the recent exchange w/ @yudapearl & @ylecun as well: https://t.co/uHqHMNU4ad. Curious how you see this in the evo-devo framing.

[Pedro Botelho @PedroBotelhoRSc]

Our NeurIPS 2025 (AI4NextG) paper is out! 🎉

Read it here: https://t.co/7y4OpZ76Lv

We adapted DQN: SNNs as function approximators.

#NeurIPS2025 #AI4NextG #MARL #SNN #Wireless #Fairness #RL #AI #NeurIPS #MultiAgentSystems #ReinforcementLearning #MultiAgentRL #DSA

[Pedro Botelho @PedroBotelhoRSc]

https://t.co/Vor5WOBFjG

[Pedro Botelho @PedroBotelhoRSc]

Neuro-cognitive radios for dynamic spectrum access

We demonstrated that Spiking Neural Networks can achieve higher fairness than established benchmarks across multiple DSA settings, while sustaining a very high throughput rate!

#neurips #rl #marl #neuromorphiccomputing

[Pedro Botelho @PedroBotelhoRSc]

[ Deep MARL air combat 🛩⚔️]

Our paper "Deep Reinforcement Learning Agents with Collective Situational Awareness for Beyond Visual Range Air Combat" is now published in IEEE Access (https://t.co/pRiJ7W8NhT)

#MARL #multiagent #reinforcementlearning #rl #ai #ml #deeplearning

[Pedro Botelho @PedroBotelhoRSc]

I am happy to share with you this new article on NU's website that features our research.

https://t.co/LKvaScAbWA

[Pedro Botelho @PedroBotelhoRSc]

Best Student Paper Award 🏆

We received the best paper student award at the 23rd WiOpt conference, in Linköping, Sweden 🇸🇪

[Trevor Gordon @todo_trev]

📡〰️〰️〰️〰️〰️〰️〰️〰️〰️〰️〰️📡
Reinforcement Learning Agents 🤝 Wireless Communication

Our paper "Fair Dynamic Spectrum Access via Fully Decentralized Multi-Agent Reinforcement Learning" just won 🥇best student paper award at WiOpt 2025. (The research started under my Masters Thesis at Columbia Under Igor Katoda and Gil Zussman and was more recently continued by Pedro Botelho, Yubo Zhang and Igor at Northwestern).

Problem: When numerous wireless devices compete for limited bandwidth, like phones📱 at a music festival, communication becomes chaotic ❌ 🛜 . Our paper presents a new approach where devices independently discover how to avoid interference and equitably share the spectrum. Key technical contributions include:

• A novel distributional Multi-Agent Reinforcement Learning (MARL) architecture using a Likelihood Hysteretic Implicit Quantile Network (LH-IQN) for decentralized cooperation 🤝. This architecture, a core part of my thesis, allows each agent to learn a distribution of possible outcomes, leading to better coordination.

• The incorporation of dynamic risk control to facilitate coordination by encouraging agents to attempt transmissions and learn effective sharing strategies 💡. In situations with limited spectrum, agents might become too cautious. Our method dynamically adjusts their willingness to take risks to ensure they actively try to transmit and discover how to share the resources efficiently.

• Fairness-driven reward structures that look at the recent history of each agent to promote equitable spectrum sharing. Instead of just rewarding successful transmissions, our reward system encourages agents to share the spectrum more evenly over time, without any central coordination.

[Networking and Internet Architecture Papers @USB]

Fair Dynamic Spectrum Access via Fully Decentralized Multi-Agent Reinforcement Learning. https://t.co/rEnl9uMz6f

[Pedro Botelho @PedroBotelhoRSc]

Our MARL paper is now available at arXiv: https://t.co/g8z2BrWhHh

Spectrum is a limited resource and the increase in its demand is already a real problem in modern society.

But what if devices could learn to share the bands adaptatively without interferences? 📶📱
#RL #MARL

[Pedro Botelho @PedroBotelhoRSc]

[New Multi-agent Reinforcement Learning paper] 🎉

Our paper proposing a novel Distributional #MARL architecture to address #DSA has been accepted at #WiOpt2025!

More details about our work:
https://t.co/sALxQiFOSB

#RL #AIAgent #ReinforcementLearning #RL #ai #ml #deeplearning