Another short nugget of writing advice that came up last week with my supervisor meetings. When you explain concepts, it’s natural to want to illustrate with an example. But watch out for the single e.g.

The single e.g. means i.e. or you are too lazy to think of more than one example.

Here are a few examples:

“There’s so much I want to do with my life, e.g. get a PhD”

“The world is changing so fast, e.g. the increasing power of the developing world”

“I love smartphones because they are so useful, e.g. you can browse the internet from almost anywhere”

What’s the problem with these statements? Their single e.g.s don’t make much sense. The sentences are trying to claim a trend (a line on a graph) when you have only given a single data point.

I’ve realised that giving one example to illustrate a concept is weak writing. If you really do want to give examples, you should not be lazy by giving only one.

Not being able to think of more than one example tells you that the stuff that comes after the e.g. is really a clarification of your point, not an illustration. In this case you should use i.e. or some variant of it.

Profiles of Office Workstation Energy Consumption Obtained Through Clustering

Clustering is one of those unsupervised learning methods that comes up time and again. I am using it heavily for behavioural grouping of people in their daily lives. The best known algorithm is called K-Means clustering. It’s quick and effective. However, there are two drawbacks to it that have prompted me to search for other methods of clustering: 1) it assumes your clusters are of the same size, and 2) you need to specify beforehand how many clusters you want.

Enter Streemer. I came across this algorithm thanks to Long Tran-Thanh, a colleague in my lab. He recommended I read the PhD of Vasileios Kandylas who was at the University of Pennsylvania and has now gone on to work for Microsoft. It is a fast, two pass (rather than iterative) algorithm for clustering, and does not assume equal cluster sizes. This is applicable to my work, because some behaviours are obviously rarer than others, and some communities are obviously smaller than others.

I implemented it in python and found some improvements in my results. The algorithm is fairly straightforward: in the first pass over the data, you do a form of streaming clustering where you cluster points that are within a certain threshold of similarity. Then you go through these (usually large numbers) of clusters a second time, you keep the clusters that have a high “cohesiveness” (i.e. high average similarity to the centroid) and integrate the rest. Finally, you find the set of outliers, the “background” cluster of points that fit into no other group.

The author states that “Streemer appears to be an ad hoc algorithm” but that there is a “connection between Streemer and online Dirichlet Process Mixture Models.” This is the part that I need to understand better. To be continued…

From the days when research contributions were more tangible

A quick tip courtesy of my supervisor. The “research contribution” is one of the most crucial sections in your progress report, thesis or paper. Without it, you are simply indulging yourself in learning knowledge on the tax payer’s dime. But it is easy to get bogged down in the specifics of how your work differs from the pack while writing this section.

One way to cut through it all is to work on the assumption that this section should contain one of these three words: first, best, only.

“This is the first algorithm that does X”
“We present the best way to do Y”
“The only work that considers Z”

The “first” and “only” cases are pretty straightforward. You’re obviously breaking new ground, and if it’s ground worth breaking then you should be honoured (for a few minutes before you have to get back to your desk).

“Best” requires more explanation. Why are you the best? It could be for several reasons: maybe your way is faster than all the other ways, uses fewer resources (memory, processing), is more reliable or robust. In short, not only is it better than everything else, but your new method is superior where it counts. This is what has to be clear in your research contributions section.

Why on earth would we want to do that?

There are many tradeoffs made in daily life that sacrifice comfort for some other benefit. Do you park further away from your destination to save on parking charges? Do you fly lowcost airlines? How about shopping around for the best price? Think about the last time you bought a chair – did you pick the most comfortable one, regardless of the cost? Most of us readily accept hardships, discomfort, interminable waiting periods, we walk further or faster, bother relatives and sales staff, all for the benefit of a few extra pounds or to help save the planet. If we can take this fact into account, it should be possible to get even more energy savings.

We already see technology flexibly capable of doing this with the “economy/green” modes in cars. Dishwashers sometimes have similar settings. I think sustainable technology in the future should do more of this. I also think it should be made one question for the user, rather than have every single piece of technology ask for it again and again. Software in your home could ask you: “do you want to put your life into green/money saving mode?” Better yet, while asking the question, we could be presented with an estimated total savings figure (in pounds or percent) for that mode. It could be a “yes” or “no” question or even a sliding scale for greater customisation. Once the decision is made, every piece of technology in our lives could be notified automatically (though the question might be made more domain specific – you might be happy to let the washing take a little longer, but want the cooking to be unaffected).

So how would this affect the experience at home or office?

A paper by Gupta et al. 2009 trialled a system that automatically lowered your thermostat while you were away from home. They used a simple algorithm that decided how quickly you could possibly get back from your current location. Since it takes time to heat a house, the heating has to be turned back on an hour before you get home (maybe), depending on the characteristics of the house and how long the heating’s been off.

There are a few ways to extend this work. The first would be to make the prediction more sophisticated. Base it on machine learning and include error ranges with each return time estimation. Once you have error ranges, you can start to be less conservative in the way you use your predictions, because the user can select the lowest accuracy level that they are prepared to tolerate. The lower the required accuracy, the greater the possible savings in energy.

Using a very basic framework and example, if the error in the estimation of your return time is 3%, then 3 times out of every 100, you will come home to a less-than-comfortably-warm house. The system originally thought you were out for the day, but it rained, so you felt like giving the park a miss.

The first thing to keep in mind is that a continuous stream of location information would allow the system to update it’s predictions. If the prediction algorithm is any good, it will notice the plan has changed quickly after you change your mind and are on your way home. This gives the heating system a little time to heat the house back up, so you don’t have to return to a stone cold house.

What would be the benefit of this small sacrifice in certainty? I don’t know, it would require investigation to find out. But the Gupta et al. 2009 paper uses ultra conservative predictions in their system and estimate the savings at 7% of HVAC energy use.

Note again that I’m not talking about flexible comfort as an expression of individual preferences in different contexts. A lot of work in the 70s looked at comfortable temperature ranges in relation to factors like the outside temperature and context (e.g. since your metabolism lowers when you sleep, you might want your bedroom to be warmer), culminating in the ASHRAE standards for building design. I would hope that any proposed heating technology takes account of personal preferences and external factors anyway. I need to learn more about the ASHRAE standards to work out how it could be taken into account.

For most of the time you would want to stick to your original decision, though occasionally you would want greater certainty, perhaps when the relatives come to stay for Christmas.

But how would people respond to a system that could automatically save them money and cut emissions? Some argue that they would use this a reason to be more profligate than usual, to the extent that any gains would be lost. If I can save you money on your heating, the argument goes, then the homeowner would just turn up the average temperature. In an analogy with transport, once EasyJet and Ryanair come on the market it becomes possible to travel much more for the same cost.

The first thing to note with such an argument is that it applies to any solution that cuts the consumption of energy, not just my particular suggestion. If you believe this argument is a fatal blow, then you also presumably think that it is pointless to reduce consumption automatically in any other way too.

I happen to believe you could start to avoid this problem if you present the choices in the right way. Especially if we add together the savings from every piece of technology in your life (i.e. if the financial saving figure is sufficiently impressive) why should we assume that the homeowner would want to pump this extra cash immediately back into the heating system? Besides, for some this could actually be a good thing: fuel poverty is a problem in the UK. A significant number of people (especially the elderly) could do with a bit more heating anyway.

The way that people respond to energy saving technology can also be dependent on the sequence of introduction. At first, the new system could be presented as a way to save money on your heating/electricity bill. This becomes an incentive to sign up. During this time an infrastructure of acceptance would be built, just as important as the physical infrastructure needed for the system. Eventually the price of energy will go up, either by government intervention or as a natural result of an upgraded (and constrained) energy supply. It then becomes the case that only the wealthy could afford to pump their gains straight back into their home energy system. Is this a dystopian picture of the future? Let me ask you, with all the perils of global warming on our doorstep, is the odd inconvenience really so unbearable?

• Ubicomp
  • Ubiquitous computing
• AAMAS
  • Agents and multi-agents
• AAAI CompSustAI
  • Sustainability using AI and computing
• CHI
  • Human computer interaction
• ICMLA
  • Machine learning and applications
• UAI
  • Uncertainty in AI

I also selected 6 academics to follow:

Prof Henry Kautz – Rochester
Prof Diane J Cook – Washington State
Prof Brian Williams – MIT
Dr Hung Bui – SRI International
Manu Gupta – MIT

Location-based behaviour analysis seemed to me like a good place to start, given the relative ease with which GPS can be used. I found 4 papers so far that talk about how energy demand can be reduced using the location of the homeowner/office worker. The general idea is that it should be possible to automatically turn off lighting and gadgets in rooms or locations where nobody is around to appreciate them. The same goes for heating when you have left the house. Some (but not all) of the work uses no predication whatsoever to make these gains. Prediction seems like the logical next step.

M. Gupta, S. S. Intille, and K. Larson. Adding gps-control to traditional thermostats: An exploration of potential energy savings and design challenges. In Pervasive ’09: Proceedings of the 7th International Conference on Pervasive Computing, pages 95–114, Berlin, Heidelberg, 2009. Springer-Verlag.

A. Roy, S. K. D. Bhaumik, A. Bhattacharya, K. Basu, D. J. Cook and S. K. Das, “Location Aware Resource Management in Smart Homes,” in Proc. IEEE Int’l Conf. on Pervasive Computing and Communications, 2003, pp. 481-488.

C. Harris, and V. Cahill, “Exploiting User Behaviour for Context-Aware Power Management,” in Proc. Wireless And Mobile Computing, Networking And Comm., vol. 4, pp. 122-130, 2005.

R. K. Harle, A. Hopper “The Potential for Location-Aware Power Management,” in Int’l Conf on Ubiquitous Computing, 2008, pp. 302-311.

The collecting of location information raises potential concerns for the user. “Will the information be private and secure?” is probably top of the list. Collecting this and other types of data may also induce the unpleasant feeling of being watched. On the other hand, people seem to be more willing than ever to share their location in return for something. We see this in the rise of Facebook Places, FourSquare, Google Latitude and many more. In return you can be presented with local coupons, be able to check-in (leading to online status), and be able to meet up more easily with your friends. According to a report done in 2010 more than 50% of people are willing to do this. Those in the 25-44 age range most happy to do so than younger or older people. If you could save on significant heating costs by sharing your location with an home management algorithm, would you do it? I probably would, though it would depend on the trustworthiness of the provider.

It is predicted that world population will be 9.2 billion by 2050 and that 16% of all people will be in the middle class by 2030 (by the United Nations and the World Bank, respectively), with most of the gains coming from currently developing nations. I believe these trends (population growth and large increases in living standards for many people), and the resulting problems with resource depletion and global warming will be the biggest factors shaping human development for the next several decades. Developing new technologies that can increase the efficiency of our energy systems and provide cleaner ways of generating electricity is a major endeavour that I think can and must take place.

I recently finished my Master’s in Artificial Intelligence at Imperial College London. The topic is highly applicable to the problem of electric grid management- in recent years there have been many interesting developments in the area of Distributed Artificial Intelligence where local actions are coordinated with local control and knowledge to give globally optimal solutions. Modelling the electric grid as a system of sensors located in each home, the problem of balancing the load (maintaining the frequency of alternating current at 50hz) can be viewed as a problem of designing agents that can make use of local information about the availability and pricing of power and act in an intelligent way, taking into account the goals of the homeowner.

Electric utilities have traditionally used homeostatic mechanisms to make the electric grid system less sensitive to variations in demand, but two major problems occur with this approach when we start to care more about reducing CO2 emissions. Firstly, the wider deployment of distributed energy resources (DER) makes it harder to provide a base load because generation is variable. Secondly, having extra production capacity to cover peak demands produces more CO2 than necessary. So the task is to introduce more decentralised intelligence into the electric grid to fill in the valleys and level off the peaks in demand and to manage the variable generation of DER. I am doing research as part of the ORCHID project and iDEaS project (Intelligent Decentralised Energy-Aware Systems) at Southampton University because I think there are several fruitful avenues to explore, and many interesting research questions that have applicability beyond the scope of this specific application area.

A key development in the electricity distribution industry that makes multi-agent systems a relevant framework for analysing and designing smart grid solutions is the widespread introduction of smart meters in homes and office units throughout the UK. Last year the government unveiled plans to put a smart meter in all 26 million homes of the UK by 2020. This opens up the possibility of adding intelligent agents to smart meters that could save money for their owners (and reduce the carbon intensity of the grid). The recent work from the iDEaS project has focused on the design of electricity markets and of agents to trade on them, and on optimal behaviours under different scenarios such as the availability of local storage. There could be plenty of further research in this area by considering how the strategy could be adapted to take account of further complexities in the model, such as users expressing preferences about the carbon intensity of the electricity they consume, or the ability to delay certain loads that are not time critical (e.g. by turning the washing machine on later, when electricity is cheaper).

Another possible area of investigation that has the potential to really improve the effectiveness of smart grid agents is the use of machine learning techniques to learn habits of homeowners and predict local future power requirements. One could analyse the patterns of consumption based upon time of day, week, month or year, but could also include information such as the weather forecast or planned cultural events. In this way, a user profile (upon which the agent would make decisions about when to buy, store or sell power) would not be a predetermined and static collection of information about the lifestyle of the inhabitants, but an adaptive and dynamic knowledge store. The sensors in this case could be modules installed in household appliances that are able to communicate with the agent (e.g. by bluetooth). During my MSc I studied several different machine learning methods and think that a Bayesian network could be used to extract patterns of behaviour from these sensors to enable the smart grid agent to predict the power requirements of the inhabitants and take steps to independently ensure that the best price is achieved.

Last year I also took an independent study option in Intention Recognition, which is the task or recognising the intentions of an agent by analysing their actions and/or the changes in the environment resulting from their actions. Usually the problem is expanded to Plan Recognition, which also recognises the likely plan that the agent will use to achieve the intention. I made a small contribution in this area by showing how the event calculus could be adapted to formulate hypotheses and their weight of evidence from observed actions in a partially observable environment. As part of the project I also wrote a limited implementation based around the plans and intentions of common daily life. Researchers in Intention Recognition often cite the concept of Smart Homes as one of the main areas of applicability (see Pereira, L.M., Anh, H.T. 2009b “Elder care via intention recognition and evolution prospection”). Smart Homes are usually focused on developing technologies to aid the elderly or cognitively impaired (e.g. Alzheimer’s patients) in the home, but what if their use could be broadened to anticipate the plans of electricity consumers in general? The very sensors that are used to monitor the demand from appliances could also be used in recognising the intentions and likely plans of the inhabitant, for which an optimal electricity trading strategy could be decided. The challenge is how to incorporate this extra information into existing trading strategies.

The transformation of the electricity market to include many more players will greatly increase the complexity of the grid but also presents opportunities for improving efficiency. A decentralised multi-agent systems approach towards optimisation under various constraints can help realise these efficiency improvements, with little or no additional effort by electricity consumers beyond installation and preference-setting. Designing agents to trade using the Adaptive-Aggressiveness strategy tailored to electricity markets and modelling their aggregate behaviour using game theory has already indicated the scale of improvement possible using this approach. I would like to investigate how the introduction of information about local patterns of consumption and user preferences could be used to achieve even better local gains and improve social welfare.