Contact

Please contact Eiko Yoneki (email: eiko.yoneki@cl.cam.ac.uk) if you are interested in any project below. * indicates high recommendation!

1.  Tuning Neural Networking Compression with Hierarchical Reinforcement Learning *

Contact: Eiko Yoneki (with Michael Schaarschmidt)

Executing neural network tasks on mobile and edge devices (e.g. facial recognition on smart phones) is gaining increasing focus in the systems community due to constraints in CPU, memory and energy usage on such devices. Consequently, neural networks trained on large servers usually need to be compressed to be fit for low power mobile execution. Recent work has explored various techniques to automatically combine various compression and quantization techniques [1, 2]. This project will investigate hierarchical reinforcement learning [3] to improve on these existing search techniques by attempting to model fine-grained dependencies between different compression and quantization methods.

[1] https://www.tik.ee.ethz.ch/file/79a7dd6f6370f809e6180c0746232283/mobisys18-liu.pdf

[2] https://arxiv.org/pdf/1806.03723.pdf

[3] https://ai.google/research/pubs/pub46646

2.  Model-based Reinforcement Learning in Computer Systems *

Contact: Eiko Yoneki (with Michael Schaarschmidt)

Reinforcement learning (RL) is gaining interest as a generic optimisation and control method in data management tasks such as resource management/scheduling, database tuning, or stream processing. A central challenge in applying RL to such problems are long decision evaluation times. A single step in a real system may take multiple minutes, as opposed to simulators for standard benchmark tasks (e.g. Atari games), which can execute thousands of steps per second.

One way of improving sample efficiency is model-based RL [1], wherein a model of the environment is learned and subsequently used to plan and evaluate decisions. A 'world-model' approach first learns a compressed version of a state representation, then a recurrent mixture density network to predict future state / reward distributions of compressed states which can be used in place of a simulator [2]. In this project, the student will evaluate a world-model based approach against a model free approach using software infrastructure built in our group [3]. Example applications include stream processing scheduling, bit-rate control in networking, static graph rewriting (e.g. TensorFlow, LLVM IRs), and various database administration tasks. The aim of the project is to evaluate different representations (e.g. graph neural networks) and planning modes to gain novel insights into cost/benefits of model-based deep RL for this domain.

[1] https://arxiv.org/abs/1803.10122

[2] https://worldmodels.github.io

[3] https://arxiv.org/abs/1810.09028

3.  Deep Reinforcement Learning in Networking **

Contact: Eiko Yoneki (with Michael Schaarschmidt)

Modern network infrastructure is increasingly driven by software-defined components. This opens up ample opportunities to explore modern data-driven machine learning methods in traffic protocols. Early work has explored deep reinforcement learning in routing [0].

Until recently, standard purpose network simulators (e.g. Ns-3 [1]) were not available via common reinforcement learning benchmark interfaces such as OpenAI gym. An open source bridge between a gym-like interface and Ns-3 has now become available [1], and several example tasks are provided. The aim of this project is to evaluate a baseline deep reinforcement learning algorithm (e.g. high-performance DQN implementation available in RLgraph [2]) on some of these tasks, and subsequently implement and explore approaches from one of the following: Hierarchical RL, model-based RL, Neural combinatorial optimization/optimal transport theory (guidance can be given depending on background and student interests).

[0] http://www.cs.huji.ac.il/~schapiram/Learning_to_Route%20(NIPS).pdf

[1] https://arxiv.org/pdf/1810.03943.pdf

[2] https://github.com/rlgraph/rlgraph

4. Optimisation of JVM Garbage Collection using ML (Bayesian Optimisation) *

Contact: Eiko Yoneki

This project extends the Cassandra case study [1,2] in incremental ways:

1)     Cassandra database was only executed in one setting. We could create an auto-tuner that is able to deal with a wider variety of settings. This includes:

-        Changing the heap size of the JVM

-        Changing the underlying hardware of Cassandra

-        Changing the underlying hardware of YCSB (Yahoo! Cloud Serving Benchmark)

2)     Applying it to other JVM applications (such as other databases, anything where we have a rigorous benchmark for throughput and latency)

3)     Looking at some of the other JVM flags (we only tuned three) as well as the other JVM garbage collector

For evaluation following two points are evaluated.

1)     No GC option as baseline

2)     Various GC implementations (i.e. some are good on certain workloads)

[1] https://www.cl.cam.ac.uk/~ey204/pubs/2017_WWW.pdf

[2] https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-900.html

5.  Dynamic Task Scheduling on Heterogeneous CPU/GPU Environment using ML for Parallel Processing

Contact: Eiko Yoneki

In this project, various aspects of parallel processing will be explored using a new generation of CPU/GPU integrated board. We use NVIDIA’s Jetson X2. This new GPUs makes it possible to cluster the GPU nodes for different scale of parallel processing. Various applications running over X2 will be investigated including graph processing and neural network training using the Tensorflow framework. Graph processing can take advantage of processor heterogeneity to adapt to structural data patterns. The overall aim of graph processing can be seen as scheduling irregular tasks to optimise data-parallel heterogeneous graph processing, by analysing the graph at runtime and dispatching graph elements to appropriate computation. Various task scheduling strategies over NN applications could be explored such as mixture of model parallelism and data parallelism.

In comparison to the above approach, several types of GPU machines in cluster computing environment will be explored. Furthermore, our recent work, BOAT [1], could be used for optimisation of complex parameter space.

[1] V. Dalibard, M. Schaarschmidt, and E. Yoneki: BOAT: Building Auto-Tuners with Structured Bayesian Optimization, WWW, 2017.  https://github.com/VDalibard/BOAT.

6.  Reinforcement Learning for Automatic Index Selection in SQL Databases  

Contact: Eiko Yoneki (with Michael Schaarschmidt)

Index selection is a standard database administration task which is often performed by human experts, typically assisted by profiling tools [0]. Selecting the optimal set of indices is difficult as it depends on query structure, data distribution, anad workload distribution. As each index incurs memory and maintenance cost when inserting new data, the aim of index selection is to identify the minimal number of indices meeting performance requirements (query latency).

Previous work in this group has implemented a framework to select compound indices in NoSQL databases such as MongoDB using deep reinforcement learning and pre-existing log data. The goal of the project is to investigate the applicability of this framework to indexing semantics [1] in relational databases such as PostgreSQL. Experimental work in the literature has explored this for single-key indices [2], but not compound cases.

In general, labelling training data is increasingly the bottleneck in deploying machine learning systems and generating labelling function based on heuristics is a growing research topic. Possibly you can extend a project for the general model building using a weak supervision [3].

[0] https://stratos.seas.harvard.edu/files/a4-idreos.pdf

[1] https://www.postgresql.org/docs/10/static/indexes-multicolumn.html

[2] https://arxiv.org/pdf/1801.05643.pdf

[3] A. Ratner, S. H. Bach, H. Ehrenberg, J. Fries, S. Wu, and C. Re. Snorkel: Rapid training data creation with weak supervision, VLDB, 2017. https://dawn.cs.stanford.edu/2017/05/08/snorkel/

7.  Optimising Neural Network Model (Hyperparameter Tuning) over Diffrent Objective Functions   

Contact: Eiko Yoneki 

In this project, you would tune neural network model itself using Bayesian optimisation based approaches [4]. Pick one typical Deep Neural Network application (e.g. image classification), and play with various hyperparameters, where hyperparameters might take substantially different values for the best accuracy or performance. Especially you can set an objective function on reducing the power consumption. You would explore various parameters based experiments and build up a probabilistic model expressing the parameter relationship applying the BOAT technique (see [1] and [2]). Also you would explore recent Google’s hyper parameter tuning on Google Cloud Platform [3], in comparison to the above approach. If time allows, you can explore Python based optimisation [5].

[1] http://www.cl.cam.ac.uk/~ey204/pubs/2017_WWW.pdf

[2] http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-900.html

8.  Optimisation of Parallel SSSP over dynamically changing road networks   

Contact: Eiko Yoneki 

The road network is static but the edge weight changes. In this project, the number of nodes in the road is 8000.  The goal is calculating SSSP for 10,000 agents, where source to destination for each agent is static). Simply calculating SSSP sequentially takes a long time. Writing multi-threading code can only increase the parallelism depending on the number of cores. The following is the basic goal and depending on how your idea evolves, it can get additional functionality.  

'Reduce the number of agents to calculate path and/or reduce the number of update on the edge weight'.  

In either case, dividing/clustering the road graph to sub-graphs. The following is a brief idea.  

a) Reduce the number of agents to calculate   -  prioritise the subgraph to recalculate based on the density of agents (i.e. <10 agents - no update of subgraph)   - adding prediction/trend of agents' movement to identify potential high density subgraphs  

b) Reduce the number of update on the edge weight   -  prioritise the subgraph to recalculate based on the volume of the edge weight update (i.e. <10 update of weight - no update of subgraph) 

9.  Building Permutation-Structured Action Spaces in Reinforcement Learning      

Contact: Eiko Yoneki  (with Michael Schaarschmidt)

In this project, you will explore indexing tasks with Reinforcement Learning. The park project has a simulator for indexing task [3] and you can see a series of indexing projects at https://github.com/park-project/park. You would apply the algorithm used in our previous work [1] to one of the park projects for transforming action space to permutation-structured action spaces. You could plug the agent code from [1] into the simulator. This will enable a scalable experiments. There is some other query processing tasks with a simulated query optimiser [2]. If time allows, you could combine it with other approaches such as graph networks.

[1] Learning Index Selection with Structured Action Spaces, arxiv https://arxiv.org/abs/1909.07440.

[2] https://github.com/park-project/park/tree/master/park/envs

[3] Park Project: https://github.com/park-project/park/tree/master/park/envs/multi_dim_index.

Contact Email

Please email to eiko.yoneki@cl.cam.ac.uk for any question.