My Projects
Stream Learning under Continual Concept Drift via Multi-Scale Model Stability Analysis
Adapting to continuously shifting data streams requires the continual dismissal of outdated data and incorporation of new data. However, when new data samples are limited, maintaining an accurate, real-time model can be quite a challenge. One possible solution is to include prior measurements in our model, but over-dependence on such historical data can introduce bias, thus impacting the model's performance. Existing methodologies often employ change detection techniques to identify the degree of reliance on prior data. However, they typically assume that data streams are piecewise stationary, changes occur during short periods of non-stationarity, and useful data exists only in segments between changes. Unfortunately, these assumptions do not hold in numerous real-world scenarios where data streams exhibit continuous concept drift, thereby reducing the effectiveness of such techniques.
In our research, we introduce the notion of 'validity horizon', representing the constantly updated optimal timeframe for which data remains beneficial for a specific prediction task. We also introduce a novel drift adaptation approach that estimates the validity horizon and evaluates the Allan variance of the learning models across multiple horizons.
This work is under review for IEEE International Conference on Data Engineering (ICDE 2024).
Adaptive Granulation: Data Reduction at the Database Level
As data volume increases rapidly, handling it effectively becomes vital. Traditional methods shrink datasets but can lose important information, especially with large and complex ones. Our new approach, Adaptive Granulation, reduces or groups data at the database level without losing crucial details. It uses a technique called Allan Variance to adjust the level of detail based on the dataset's structure, ensuring a balance between data reduction and prediction accuracy. This work is accepted for KMIS 2024 conference.
Detection of Surface Friction Conditions from a Fleet of Vehicles (Ongoing project)
In the field of intelligent transportation systems, connected vehicles may benefit from sharing roadway conditions experienced by neighboring vehicles, especially in safety-critical scenarios such as the presence of black ice ahead.
Such safety-critical scenarios can benefit significantly from data collection and storage systems that are intentionally forgetful by design, such that data is retrieved and shared in safety-critical situations only within the time and location regions for which they are relevant. As shown in the first figure, we consider fleets of connected and autonomous vehicles (CAVs) driving on potentially icy roads, where safety-critical road friction information is shared via a wireless data link to a central database that mediates data averaging. In this project, the database is considered as a sensor for the vehicle, and the aggregation of data within the database as a sensor-averaging process. The averaging operation, however, needs to be carefully performed such that only recent and valid information is used to build a spatial friction model that is updated on-the-fly as new vehicular information is received.
Visit The official project webpage for more information.
Reward-Sharing Relational Networks in Multi-Agent Reinforcement Learning as a Framework for Emergent Behavior
Most prior works on MARL seek to implement intra-agent complex interactions by explicitly communicating agent actions. However, there have only been a few efforts that examine emergence as arising from complex ‘social’ interactions or relations based on individual objectives and reward functions. This study is about integrating a user-defined relational network into the MARL setup and evaluating the effects of agent-agent relations on the generation of emergent behaviors. Specifically, we propose a framework that uses the notion of Reward-Sharing Relational Networks (RSRN) to determine the relationship between agents where edge weights determine how much one agent is invested in the success of (or ‘cares about’) another. The preliminary results indicate that reward sharing relational networks can effectively influence the learned behaviors towards the imposed relational network.
You can also check this webpage for more information
RSRN_Demo.mp43 agents try to cover 3 landmarks while agent 3 (yellow) is systematically slowed down by limiting its possible actions.
Near-optimal Moving Average Estimation at Characteristic Timescales: An Allan Variance Approach
A major challenge in moving average (MA) estimation is the selection of an appropriate averaging window length or timescale over which measurements remain relevant to the estimation task. Prior works typically perform timescale selection by examining multiple window lengths (or models) before selecting the `optimal' one using heuristics, domain knowledge expertise, goodness-of-fit, or information criterion (e.g., AIC, BIC etc.). In the presented work, we propose an alternative mechanism based on Allan Variance (AVAR) that obviates the need for assessing multiple models and systematically reduces reliance on heuristics or rules-of-thumb. The Allan Variance approach is used to identify the timescale that minimizes bias, thus determining the timescale over which past information remains most relevant.
This work has been presented at the American Control Conference (ACC) 2021.
Thermodynamics-inspired Modeling of Macroscopic Swarm States
The collective behavior of swarms is extremely difficult to estimate or predict, even when the local agent rules are known and simple. But what comes before this, is quantifying of the collective behavior. The presented work seeks to leverage the similarities between fluids and swarm systems to generate a thermodynamics-inspired characterization of the collective behavior of robotic swarms. This work has been presented in the Dynamic Systems and Control Conference (DSCC 2019) .
Prey and Predator Swarm Simulation
In this project, a swarm is system simulated along with some attracting resources and threatening predators. The swarm dynamic is modeled based on Couzin model (Couzin et al., 2002).
Migration Control in Nonlinear Dynamic Systems
Limit cycles appear in many multi-dimensional dynamic systems. For systems with more than one limit cycle, it might be possible to change the current phase (phase migration) by altering a specific parameter in the system. In this work, I identified the minimum time interval required for such migration.
Real-time Detection of Swarming Spheros
My primary objective in this project was to develop an experimental platform to delve into swarm systems, using Sphero robots as the primary agents. Throughout this initiative, my main concentration was on detecting these robots via image data and subsequently providing both positional and orientational feedback to the swarm behavior controller. In the detection phase, I initially contemplated utilizing the LED colors of the Spheros for identification. However, recognizing the potential for color overlaps, I pivoted to a cluster-based approach. After capturing a Sphero's color and transitioning the image from RGB to HSV, I observed a notable enhancement in detection quality. Building on this, I employed the k-means clustering algorithm to ascertain each Sphero's precise location. This methodology was further refined when I projected a Sphero's subsequent position based on its current one, using this data to search for the nearest cluster and accurately label each robot.
Addressing the challenge of noise interference in image processing, I introduced a two-pronged strategy. Firstly, I designated a "valid zone" for each agent, effectively countering false detections caused by noise. Secondly, I employed image filtering techniques to further diminish noise-related anomalies. I then translated positions identified on the 2D image plane into 3D real space using triangulation and designated landmarks on the ground. In addition to these advancements, I meticulously formulated methods to determine the robots' headings, emphasizing the criticality of discerning not only a Sphero's location but also its intended direction.
Adrina: an Autonomous Search & Rescue Robot
Adrina is a 4-wheel steered autonomous robot which is designed for Search and Rescue operations. It has been designed and developed at "Advanced Mobile Robotics Lab" of Qazvin Azad University (QIAU). As a mechanical engineer, I was responsible to design, test and evaluate a wheeled mobile robot platform capable of maneuvering on uneven and rough surfaces. It had to be able to move and steer along curvature paths without any slippage.
Our team, MRL, participated in two international RoboCup rescue competitions in 2014 and 2015 with Adrina. We won the second place in Robucup 2014 and the first place in RoboCup 2015.
