Research
My research sits at the intersection of automated negotiation, multiagent systems, robotics, and machine learning. Automated negotiation has been the connecting thread for the last decade; before that, my PhD and postdoctoral work focused on human-robot interaction and the time-series/pattern-discovery tools needed to make sense of embodied interaction data — tools that I still use today when mining negotiation logs and sensor streams. The four areas below give an overview, each with a handful of representative publications (the full list and patents have the complete record).
Automated Negotiation
This is the focus of most of my work since 2018. Automated negotiation asks how autonomous agents can reach agreements on their own — over contracts, resources, or schedules — without a human in the loop for every deal. I work on the theory (protocols, optimal strategies, and mechanism design) and on learning-based negotiating agents.
On the theory side, I introduced the TAU (Tentative Acceptance Unique Offers) protocol, a negotiation protocol that lets agents reach efficient agreements without ever revealing — or even needing to model — their partner’s utility function (Mohammad, 2026, Mohammad, ECAI 2025). This builds on earlier work on optimal time-based bidding strategies (Mohammad, 2023), generalized bargaining protocols that unify auctions and negotiations under one framework (Mohammad, AJCAI 2023), and concurrent negotiation — where an agent runs many linked negotiations at once against a single global utility function (Mohammad, 2021). I also work on preference elicitation during the negotiation itself using a Value-of-Information criterion so an agent only asks for information that is worth the cost of asking (Mohammad & Nakadai, AAMAS 2019).
Reinforcement learning is a growing part of this agenda: with Ayan Sengupta we built RL-based negotiating agents that adapt their strategy mid-negotiation and transfer across domains and opponents (Sengupta, Mohammad & Nakadai, AAMAS 2021, Sengupta, Nakadai & Mohammad, IJCAI 2022), and I have argued for treating negotiation as a distinctive middle ground between single-agent RL and full general-game-play for the MARL community. This line of work is now moving toward using generative AI — LLMs as negotiators and as tools for negotiation analytics (Mohammad, Chen, Higa, Ando & Morinaga, 2025) — and toward using time-series forecasting to make procurement negotiation proactive rather than reactive (Mohammad & Chen, AAMAS 2026). Most of this work (10 patents and counting) is done jointly with NEC Corporation, where it feeds directly into industrial negotiation-support products.
Multiagent Systems
Automated negotiation only becomes useful at scale when many agents interact, so a large part of my work is the multiagent infrastructure and benchmarks that make that study possible. Much of this research would not be reproducible without shared infrastructure, so I built and maintain NegMAS, an open-source platform for developing and evaluating negotiating agents embedded in simulations (Mohammad et al., PRIMA 2020).
On top of it I created the Supply Chain Management League (SCML), an ANAC competition track and benchmark environment where agents negotiate contracts up and down a multi-tier manufacturing chain — from raw silicon to printed circuit boards in the running example — under real-world pressures like bankruptcy risk and delivery penalties (Mohammad et al., JSAI 2020, Mohammad et al., PRIMA 2019). More recently, SCML was repackaged as a generalist Gymnasium/PettingZoo environment for RL and MARL research (Mohammad, Nakadai & Greenwald, PRIMA 2024).
I co-organize the Automated Negotiation League (ANL) and the Human-Agent Negotiation (HAN) league at ANAC, and have co-authored the annual ANAC challenge/results reports for several years running (2025, 2024, 2019).
Robotics & HRI
My PhD (Kyoto University, under Toyoaki Nishida) and the years after it were spent asking a simple question: how can a robot learn to interact naturally with people, the way people learn to interact with each other — by watching, imitating, and adapting — rather than through hand-coded scripts? This work is summarized in the book Data Mining for Social Robotics: Toward Autonomously Social Robots (Mohammad & Nishida, Springer, 2015).
A recurring theme is imitation as a discovery problem rather than a copying problem: before a robot can imitate a demonstrator, it has to work out what is worth imitating out of a continuous, noisy stream of motion. My “Fluid Imitation” framework addressed exactly this (Mohammad & Nishida, 2012), and later work looked at how humans judge a robot’s imitative skill, including the surprising finding that a robot imitating the human back (back-imitation) changes how natural its behavior is perceived to be (Mohammad & Nishida, 2015). Related papers tackled the correspondence problem for mapping human gestures onto a humanoid’s very different body in closed form (Mohammad & Nishida, AMT 2013), and generating human-like motion for a humanoid performing a shadowing task (Mohammad & Nishida, CTS 2014). With Mostafa Hussein, this line continued into practical learning-from-demonstration tools, including a ROS package for continuously teaching a robot arm to write by demonstration (Hussein, Mohammad, Ali & Nishida, ICMA 2017) and a variational-Bayesian approach to LfD (Hussein, Mohammad & Ali, IEA/AIE 2015).
The other half of natural interaction is perception and attention: I worked on autonomous development of gaze behavior so a robot can learn where and when to look during a conversation instead of using scripted gaze patterns (Mohammad & Nishida, 2013, Mohammad & Nishida, 2010), and on using physiological signals (skin conductance, heart rate) as an objective proxy for how “natural” an interaction felt to the human participant (Mohammad & Nishida, 2010). More recently, with Abdullah Ahmed and colleagues, this extended to end-to-end deep reinforcement learning for mobile robot navigation in dynamic environments shared with people (Ahmed, Mohammad, Parque & El-Hussieny, MESA 2022).
ML & Time-Series
Understanding the interaction data behind the robotics work required tools that did not exist yet at the time, which led to a parallel research thread on unsupervised time-series analysis: finding when a multidimensional signal changes behavior (change-point discovery) and which sub-sequences repeat meaningfully across it (motif discovery), without knowing in advance what to look for or how long a pattern should be.
The starting point was Robust Singular Spectrum Transform (RSST), a change-point detection method robust to noise and outliers that won a Best Paper Award at IEA/AIE 2009 (Mohammad & Nishida, 2009), later extended and benchmarked against other SSA-based algorithms, again with a best-paper recognition (Mohammad & Nishida, SII 2011). On top of change-point discovery I developed a family of constrained motif discovery algorithms that find approximately-recurring patterns whose length and number of repetitions are not known beforehand (Mohammad & Nishida, 2009), including exact algorithms that are simultaneously invariant to the pattern’s length, scale and mean (Mohammad & Nishida, 2016). These were packaged into two open-source Matlab toolboxes, CPMD (Mohammad, Ohmoto & Nishida, IEA/AIE 2012) and its successor MC², which adds causality discovery between processes to the same pipeline (Mohammad & Nishida, IEA/AIE 2016).
On the broader machine-learning side, I have a long-running collaboration with Panikos Heracleous and colleagues at KDDI Research / NEC on deep learning for speech and sensor signals — mostly speech emotion recognition and spoken language identification using deep CNNs, i-vectors, and more recently GANs and Vision Transformers (Heracleous, Mohammad & Yoneyama, HCII 2020, Heracleous, Fukayama, Ogata & Mohammad, HCII 2022), and a hierarchical federated GAN, HFL-GAN, that scales to many heterogeneous clients without centralizing their data (Petch, Moustafa, Ma & Mohammad, 2025). A second thread, from my time at KDDI Research, is human activity recognition from wearable sensors: orientation-insensitive features (Mohammad, Matsumoto & Hoashi, 2019), recognizing “primitive” activities from very short sensory sequences (Mohammad, Matsumoto & Hoashi, 2018), and a public dataset of smart-watch accelerometer recordings from an unmodified kitchen (Mohammad, Matsumoto & Hoashi, MUM 2017). The same time-series machinery originally built to mine human-human interaction records now underpins how I analyze negotiation logs and sensor streams today.