Danielle Varjosalmi

I am a PhD student at the Mathematics for Real-World Systems CDT.

My areas of academic interest include: simulation-based optimisation, machine learning, and game theory. In particular, I am interested in the simulated annealing and Bayesian optimisation algorithms.

Current Work

My ongoing project involves exploring techniques to parallelise the simulated annealing algorithm for implementation in both cloud-based architecture and smaller multicore systems. I focus on efficient parallelisation of runs in which the objective function is discrete, expensive and stochastic.

The project is in collaboration with Twinn (formerly Lanner Ltd), a company that produces simulation software to model complex, stochastic processes. I am supervised by Prof. Juergen Branke (Warwick Business School) and Prof. Robin C. Ball (Physics).

Speculative Computation

Speculative computation is the use of idle processing power to speed up the completion of an uncertain sequence of tasks by gambling on potential tasks in the sequence before they can be known to be in it for certain. The success of the parallelisation depends on the success of the gambling decisions.

Optimal Strategies

We implement a method for simulated annealing by formulating the resource allocation problem of cores to samples as a Markov decision process. This allows us to combine the steady state with a payoff matrix to quantify the performance of individual strategies. We use dynamic programming to find the optimal strategy given an acceptance rate.

Optimal Execution

We propose a dynamic estimation method for the acceptance rate using Bayesian heuristics and explore its use to guide effective speculation. The performance of the estimator and the behaviour of the system using different scheduling protocols are evaluated on a specialised forward simulator we built.

Parallel Tempering

Parallel tempering is an algorithm that condenses the sequential annealing schedule used in simulated annealing into multiple fixed temperature chains that are run in parallel but interact through occasional swap tests.

Realistic Analysis

We test variations of this technique comprehensively on a realistic simulation benchmark provided by Twinn and against both Twinn’s commercially available WITNESS Optimiser algorithm and Bayesian optimisation.

Academic Background

Warwick Business School, MSc Business Analytics with Distinction (Top 10%)
University of British Columbia, BA Economics
University College London, Exchange

Summer Schools:

2019: Gaussian Process and Uncertainty Quantification Summer School (GPSS)
2019: Complex Networks: Dynamics and Control (MathSys and DSSC)

Additional Activities

Reading Groups:

I started and am actively involved in the Bayesian Optimisation Reading Group that meets virtually every Wednesday at 10:30 on MS Teams. We discuss significant optimisation papers, present our latest research, collaborate on relevant projects, and participate in competitions. Our group includes researchers from MathSys, WBS, and international collaborators. See the link above for more information and get in contact to be added to the group!

Teaching:

2020/21: Seminar Tutor for IB9V60 Discrete Event Simulation at WBS
2018/19, 2019/20, 2020/21: Seminar Tutor for IB2110/IB3200 Simulation at WBS

Other Positions:

2018 - Present: Centre for Complexity Science Newsletter Editorial Staff
2017 - 2018: Enactus Warwick Consulting Team Leader
2017 - 2018: MSBA SSLC Secretary

Contact Details

d.varjosalmi@warwick.ac.uk

Office: D2.05

Complexity Science,
Zeeman Building,
University of Warwick,
Coventry,
CV4 7AL