Wissal Bekhti

Research

Cooperation among learning agents

My doctoral research studies how cooperation emerges — or fails to emerge — among self-interested agents trained with reinforcement learning. In semi-cooperative settings, agents share an environment but not a common objective; my work designs incentive structures, such as reputation mechanisms and filtered reward shaping, that align individual goals with collective outcomes in high-dimensional environments.

Doctoral research

Reputation mechanisms for semi-cooperative multi-agent RL

2024 — present

AI Movement, Mohammed VI Polytechnic University (UM6P)

Starting from a semi-cooperative prey–predator game, this line of work introduces a reputation mechanism and filtered reward-shaping functions that encourage agents with heterogeneous abilities to cooperate. Our approach is evaluated against established baselines such as IQL, Partial Q-MIX, and PED-DQN, and was published at AAMAS 2025 with an application to collaborative target acquisition and coverage optimization in satellite constellations.

MARLReward shapingReputationSatellite constellations

Selected projects

Embodying a qubit's mind with stochastic deep RL and quantum computing

2022 — 2023

ENSAM Meknès

A mathematical framework for studying qubits — Hilbert spaces, density matrices, Bell states, and reversibility under entanglement — combined with a Q-learning agent. Real measurement data were collected from IBM Quantum Lab, with circuits built in Qiskit, to predict a qubit's next spin with a classification model.

Quantum computingQiskitQ-learning

Teaching quantum agents to stay entangled inside a wormhole

2023

Independent project

An exploration of the ER = EPR conjecture through the lens of reinforcement learning: studying the entanglement of qubits on opposite sides of an Einstein–Rosen bridge, information flow at the event horizon, and a QOMDP-based RL model simulating the experiment on a holographic wormhole.

Quantum RLQOMDPEntanglement

Time-complexity optimization for Branch & Bound via machine learning

2023

ENSAM Meknès

On the Travelling Salesman Problem, K-means clustering of scattered instances is used to reduce the effective time complexity of the Branch & Bound algorithm, with a systematic comparison of the clustered and exact approaches.

Combinatorial optimizationTSPClustering

Optimizing Q-learning on Atari games

2023

ENSAM Meknès

Improvements to Q-learning agents on Pong and Breakout through a target network and learning-rate scheduling, benchmarked against DQN and Double DQN.

Deep RLDQNAtari

Sun-tracking model for an intelligent energy cell

2023 — 2024

ENSAM Meknès × UQAR (Canada)

A custom Gym environment models a solar energy cell; reinforcement learning policies are trained to maximize captured solar energy under different conditions.

RLGymRenewable energy

AI-driven workforce schedule optimizers

2023 — 2024

ENSAM Meknès

A customized scheduling problem solved with genetic algorithms, the Hungarian algorithm, linear programming, and SVMs on generated data — delivered as a Streamlit app displaying smart schedules alongside statistical analyses of productivity and fatigue.

SchedulingGenetic algorithmsStreamlit

Smart electronic voting with blockchain and big data

2023 — 2024

ENSAM Meknès

Election data stored off-chain with IPFS, a Solidity smart contract for vote counting tested on Tenderly and validated on Ethereum, and analysis of validated votes with Apache Spark and MapReduce.

BlockchainSolidityApache Spark

Twin and cousin primes in the forming of all even numbers

2022

Independent project

Building on Goldbach's and Polignac's conjectures, this project proposes two new conjectures relating twin and cousin primes to the decomposition of even numbers, with algorithms for checking their validity and a study of the correlations between them.

Number theoryConjectures

Industry experience

Deep RL for satellite constellations — research intern

Feb — Jul 2024

AI Movement (UM6P), Rabat

Collaborative target acquisition and coverage optimization applied to a satellite constellation; the internship that grew into my current doctoral research and our AAMAS 2025 paper.

Quality control for DAP fertilizer manufacturing — intern

Aug — Sep 2023

OCP Group (DAP/JFC4), El Jadida

Six Sigma methodology applied to enhance a control system, real-time Power BI dashboards fed by push datasets from SQL Server, and an interactive Flask interface with embedded reports providing corrective recommendations to operators in case of anomalies.

Predictive maintenance with recurrent neural networks — intern

Jul — Aug 2022

Nestlé Factory, El Jadida

An LSTM pipeline for sensor data — collection, smoothing, correlation analysis, and dimensionality reduction — with model evaluation and a study of time and space complexity.