Projets novateurs réalisés

Deep Reinforcement Learning in Optimal Market-Making

On June 1, 2021, Futures First Canada and FinML began a pilot collaborative project involving three Canadian universities by-way-of a MITACS Accelerate internship (IT25712) which jumpstarted an initiative to use cutting edge techniques in machine learning, financial mathematics, and AI for making predictions in financial markets. This goal is integral to the business operations of Futures First and is currently a popular topic in academic research. The pilot project created a basis for this current project application which will continue the effort, extending positive academic results and furthering integration into the company’s infrastructure. This next project will extend previous research from project IT35694 which focused on applying SOC (stochastic optimal control) to algorithmic and HFT (High-Frequency Trading) problems. The trading problem Futures First specifically focuses on is market-making, which is one of the most heavily studied trading problems in the futures trading environment. The main subject of this new project will be to utilize the novel topic of deep reinforcement learning to generate optimal market-making solutions. This will be a challenging project where we uncover the obstacles for applying the academic theory around deep reinforcement learning in practice.

Procedural Tree Modeling with Silhouette Constraints

Modeling trees is known to be a difficult problem in computer graphics. At Animal Logic, the typical artist workflow involves using the in-house tool for semi-procedural tree generation. The artists hand model trunk, branch, and leaf geometry, and trees are generated by scattering branches and leaves in a recursive manner.

This approach has several limitations. First, artists often would like to hit a certain silhouette or shape of a tree, represented with closed mesh volumes. While this is possible with the existing approach by culling branches that leave the volumes, this results in an unnatural appearance, as if the tree had been manually trimmed. Second, the approach does not prevent intersections between branches during the recursive addition of branches. Such cases are undesirable and result in an unrealistic appearance of generated trees.

The aim of this project is to develop a method for semi-procedural tree generation that generates a tree defined within mesh volumes using artist modeled trunk, branch, and leaf geometry in a manner consistent with artist expectations. Our approach will significantly reduce the time artists spend manually creating trees that fit a certain shape; as a result, large collections of trees can be constructed with reduced manual intervention.

Development of Cost-Effective Solutions for Energy Management in Smart Buildings

The proposed research aims at developing control strategies under the paradigm of Demand Response (DR) in the context of the Smart Grid in order to improve energy efficiency and to reduce operational cost in commercial buildings and communities. The emphasis will be put on consumer side energy management strategies that able to balance energy demand and supply and to reduce the overall operational cost while providing an enhanced performance. The envisaged solutions lie mainly on autonomous demand response management in smart buildings including peak shaving, consumption scheduling, and load forecasting. The achievements of the present project will allow the industrial partner, Fusion Energy Inc., to enhance their solutions for energy management through optimization of mechanical and electrical equipment, automation, and real-time energy consumption control.

Micronuclei detection using immunofluorescence images

“THIS IS A GENERIC TEXT PUT IN PLACE AS THERE WAS NO PROJECT OVERVIEW”

Machine Learning-Driven Decision Support for Autonomous Services in Airport Operations

Centered on airport operations, this research utilizes open flight data from multiple airports as a focal point. Collaborating with Aurrigo, the project aims to optimize data acquisition from open sources and construct, train, and thoroughly assess machine learning models for forecasting future airport operations. These predictive capabilities will play a pivotal role in informing decision-making processes regarding the deployment and strategic planning of autonomous services within airport facilities. By integrating advanced predictive analytics powered by machine learning, this project aims to transform the operational efficiency of autonomous services, facilitating cost-effective and strategically informed decision-making across various operational domains.

Pioneering Digital Organs for Next-Gen Translational Medicine

“THIS IS A GENERIC TEXT PUT IN PLACE AS THERE WAS NO PROJECT OVERVIEW”

Causal Discovery from Non-Stationary Time Series

“THIS IS A GENERIC TEXT PUT IN PLACE AS THERE WAS NO PROJECT OVERVIEW”

Developing a novel PPP-RTK location accuracy technique for the Android platform

This project is about developing a new technology for Android smartphones that improves their ability to determine exact locations using satellites. This is challenging due to issues like atmospheric effects and currently there’s a gap in research for this technology in the Android market. The plan is to use advanced methods that are in line with global standards to make the GPS positioning on Android phones much more precise. This could be a major breakthrough, setting new standards in the industry and making it easier for people around the world to use location-based services accurately. The goal is to create an App for Android phones that can pinpoint locations very precisely, using only the phones itself, not needing any extra devices.

In vivo characterisation of a newly engineered D-serine biosensor with a micro-optrode

The brain’s 100 billion neurons communicate through the release of neurotransmitters and neuromodulators across small junctions (synapses). The last two decades have seen remarkable advances in understanding the role of such molecules in the nervous system. Amino acids such as D-serine, are now recognized as a vital neuromodulators for synaptic plasticity but also for their involvement in many pathologies. Accordingly, D-serine signalling supports long term changes in synaptic plasticity and cognitive performances while signalling aberrations have been consistently associated with several pathological
conditions including schizophrenia, Alzheimer’s disease and epilepsy. Despite these observations, we still lack a thorough understanding of how brain activity influence variations in D-serine and the mechanisms by which this amino acid impacts brain synpases. This project will validate the use of a new genetically encoded light sensitive biosensor able to detect D-Serine in the intact rodent brain. It will improve our understanding of the D-Serine mechanisms and lead to new strategies to develop therapeutics for brain diseases.

Assessment of Machine Learning–Based Medical Directives in Pediatric Emergency Medicine

“THIS IS A GENERIC TEXT PUT IN PLACE AS THERE WAS NO PROJECT OVERVIEW”

Neural Network Model for Predicting NBA Shot Outcome

As the game of basketball evolved, analysis of the game has also grown from taking average of field goal percentage to more complex analytics. In the 2013-2014 season, the NBA has installed the SportVU Player Tracking technology in every NBA arena. SportVU collects 25 frames of data per second, each frame containing the (x,y) coordinates of each of the 10 players and the (x,y,z) coordinates of the basketball. The goal of this research is to understand how much better we can predict the outcome of shot given this massive amount of newly available information, and what the important factors are in contributing to a made shot. This understanding will assist the Toronto Raptors, and even the general basketball community, on many levels. For example, this could provide some guidance to players (shot location and time-of-game selection) and to coaches (which player/game situations tend to be most successful. The system could also help quantify the quality of performances of players by the shots that they took, instead of only looking at the outcome, which is inherently probabilistic.

Conversational Problem Solving: generating multi-step actionable plans to build trustworthy dialogue agents

THIS IS A GENERIC TEXT PUT IN PLACE AS THERE WAS NO PROJECT OVERVIEW