Professor of Machine Learning at the Department of Computer Science at Royal Holloway, University of London.
My research interests include reinforcement learning, evolution and evolutionary algorithms, immunology, kernel methods in machine learning, LLMs and reading, and visualisation of high dimensional data
I connected RL to Markov decision processes and invented Q-learning in the late 1980s in my unpublished PhD thesis. For the story of this, please see Reinforcement Learning: some history.
My main current research interest is evolution and evolutionary algorithms. My motivation is that animals have evolved to learn fast, because they do not have time to learn slowly. Evolution is the primary algorithm: how is it so effective?
One starting point is to ask how complex can organisms in principle become? How can we define the complexity of an organism? Are certain types of genetic codes (codes in a general sense – not just triplet codes) more evolutionarily efficient than others? These are rather basic questions in theoretical biology, and a new approach to answering them is outlined in Selective Breeding Analysed as a Communication Channel, Watkins (2008).
The trick is to define communication channel(s) based on selective breeding: the capacities of suitably chosen channels can be calculated, or estimated from simulations, and these provide strong upper bounds on possible organismal complexity. This approach can be extended to consider the relative efficiencies of different schemes of genetic encoding, and to provide insight into the adaptive complexities of communities of organisms.
Next, what is the most appropriate computational abstraction of evolution? We have been developing MCMC quasi-Bayesian evolutionary algorithms, which allow evolution with recombination to be modelled as Gibbs sampling, producing a Markov chain of populations that satisfies detailed balance. See Lember and Watkins (2022), and Poulton, Altenberg and Watkins (2023).
An early paper on kernel methods in which I assisted is Weston and Watkins (1999), which presents Jason Weston's method of formulating the support vector machine optimisation to distinguish multiple classes.
After that I became interested in whether kernel methods could be applied to non-vectorial objects, such as strings of different lengths. Watkins (1999) introduced a class of kernels that can be applied to strings. Other people also had similar and perhaps better ideas: David Haussler, Tommi Jaakola, and Alexei Chervonenkis.
String kernels were used in Lodhi et al (2002), which was perhaps the first paper that presented positive results from learning with string kernels. This work helped to stimulate considerable subsequent research on kernels of this and similar types.
My colleague Alex Clark then suggested that string kernels might be used to define "planar languages", which are languages in which all sentences lie on a hyperplane in a feature space induced by a string kernel. Clark et al (2006) and Clark et al (2010) present results in this area.
I collaborated with Prof Vincent Jansen in at Royal Holloway in a theoretical project on the spread of epidemics. The aim was to study how an epidemic might interact with the spread of information about the epidemic, which itself would alter people's behaviour. One possible model of this is presented in Funk et al (2009). I was proud to have a small role in this work.
In 1996, Robert Macrae (now at the Systemic Risk Centre at LSE) and I considered the problem of estimating the price volatility of investment portfolios, and of using these estimates in portfolio optimisation (Macrae and Watkins 1996). To our surprise, we found that standard investment textbooks recommended optimising portfolios using entirely unsuitable risk estimators. Indeed, the textbook methods of portfolio optimisation – which were also widely sold by major investment banks – could recommend meaningless portfolios and cause serious overtrading, which was of course to the banks' advantage.
The standard method was to use the empirical covariance matrix of price changes as the covariance estimate for optimisation of a portfolio. The problem with this is that portfolio optimisation requires an estimate of the inverse of the covariance matrix, which is also known as the precision matrix — but simply inverting the empirical covariance matrix gives a terrible estimate of the precision matrix. We recommended simple and practical ways of avoiding this error.
Our papers were never published, although a short summary of our results did appear in Macrae and Watkins (1999). With 25 years of hindsight, it appears that we were among the first people to write about this problem, and elements of our research turned out to apply to the problems of both LTCM and the 2008 financial crisis.
Selective Breeding Analysed as a Communication Channel
C.Watkins, SYNASC 2008 (Timisoara)
An Evolutionary Model that Satisfies Detailed Balance
J. Lember and C. Watkins
Methodology and Computing in Applied Probability 2022
https://link.springer.com/article/10.1007/s11009-020-09835-5
Evolution with Recombination as Gibbs Sampling
J. Poulton, L. Altenberg, and C. Watkins
Theoretical Population Biology 2023
https://doi.org/10.1016/j.tpb.2023.03.005
Languages as Hyperplanes: Grammatical Inference with String Kernels
A. Clark, C. Costa Florencio, C. Watkins
Machine Learning (2010)
Published online http://dx.doi.org/10.1007/s10994-010-5218-3
The Spread of Awareness and its Impact on Epidemic Outbreaks
S. Funk, E. Gilad, C. Watkins, and V. Jansen
Proceedings of the National Academy of Sciences 106 (2009)
Languages as Hyperplanes: grammatical inference with string kernels
A. Clark, C. Costa Florencio, and C. Watkins
European Conference on Machine Learning 2006 (Berlin)
This paper won the ECML Innovative Contribution Award
Text Classification using String Kernels
H. Lodhi, C. Saunders, J. Shawe-Taylor, N. Cristianini, C. Watkins
Journal of Machine Learning Research 2 2002 (Feb) : pp419-444
Dynamic Alignment Kernels
C. Watkins,
In A.J. Smola, P. Bartlett, B. Scholkopf, and D Schuurmans (eds)
Advances in Large Margin Classifiers, MIT Press 1999
Multi-Class Support Vector Machines
J. Weston and C. Watkins,
6th European Symposium on Artficial Neural Networks (ESANN 99)
Safe Portfolio Optimisation
Robert Macrae and Chris Watkins
IX Symposium on Applied Stochastic Models and Data Analysis (ASMDA), Lisbon, 1999
A Disaster Waiting to Happen [pdf]
Robert Macrae and Chris Watkins,
a working document from 1996, unpublished.