Agent ≅ Human

I have been trying to read the articles from the most recent conferences and journals, and I firstly started to check papers in AAAI 2012. There are two articles that I found interesting and related. The articles are as follows.

- Strategic Advice Provision in Repeated Human-Agent Interaction

The paper assumes that agents using equilibrium strategies to interact with human are not successful in repeated settings. It asserts the idea that an agent considering only its own outcome while providing suggestions cannot perform well in long term as the user will learn to deny its suggestions. Hence, the suggestion should be based on the weighted sum of both the agent and the user outcomes. The idea is tested through an experiment. The experiment examines the interaction between a navigation system and a driver. The navigation system (the agent) aims to minimize fuel consumption, while the driver's goal is to reduce travel time. The designed agent which utilises a social utility approach in which both fuel consumption and travel time are tried to be reduced reasonably can gain more  through increasing the user's trust in long term.

 

* The similarity: we aim to form an agent flattening the peak electricity demand, while consumers want to maintain their comfort levels.

 

- Agent-Human Coordination with Communication Costs under Uncertainty

 

In this paper an agent design is proposed for efficient agent-human coordination in settings where uncertainty and incomplete information exist. Simply, the agent predicts the people's behaviour through neural network and decides which specific information to deliver to teammate based on the communication cost and the possible outcomes of the information.

 

* As we are going to use messaging agents among energy consumers, it will be necessary to take the communication costs and the possible outcomes into account so as to increase performance and efficiency.

 

 

I read the following papers which are cited in the papers stated above.

 

- A Study of Computational and Human Strategies in Revelation Games

- Computers in Human Behaviour

- Using Focal Point Learning to Improve Tactic Coordination in Human-Machine Interaction

- Credibility and Determinism in a Game of Persuasion

 

 

Team Messaging

This week I added messaging feature to FigureEnergy. Hereafter team members are able to contact each other through the new feature.

I also read a bunch of academic papers:

- SmartCap: Flattening Peak Electricity Demand in Smart Homes

- Agent-Based Control for Decentralised Demand Side Management in the Smart Grid

- Multi-Agent Control of Thermal Systems in Buildings

- Smart Heat Grid on a Intraday Power Market

- An Agent Framework for Knowledge-Based Homes

- Improving Building Energy Efficiency with a Network of Sensing, Learning and Prediction Agents

- Intelligent Heating Systems in Households for Smart Grid Applications

- Two Theories of Home Heat Control

- An Intelligent Agent for Home Heating Management

- A Scalable Low-Cost Solution to Provide Personalized Home Heating Advice to Households

I will update this post with providing short descriptions of the papers. Meanwhile, I am reading  papers related to energy from top conferences and journals. I also started to read the book 'Pattern Recognition and Machine Learning' by Bishop.

Human-Agent Interaction

This post will summarise  TeamCore research group's studies which are related to energy efficiency and human-agent interaction. http://teamcore.usc.edu/junyounk/energy/

- Human-Building Interaction for Energy Conservation in Office Building

The study aims to increase energy awareness through a reminder system which delivers simple and complex messages to occupants in order to change their energy consumption behaviour. While simple messages only involve suggestions about how occupants can behave to decrease energy consumption, complex messages additionally include information about the consequences of the suggested behaviours. The point emphasized in the study is that intensive feedbacks that contain the outcomes of the occupants' behaviour are more powerful than simple suggestions to convince people to adopt green habits.

 

- Towards Robust Multi-Objective Optimization Under Model Uncertainty for Energy Conservation

The aim of the work is to maximise energy savings without affecting the comfort level of occupants in commercial buildings. There are two types of agents which are room agents and proxy agents. Proxy agents represent occupants through containing occupant's model, and they interact with Room agents on behalf of occupants. Room agents represent the offices or conference rooms, and their task is to reduce energy consumption in the represented room through negotiating with Proxy agents to minimise the use of office devices (lights, HVAC, etc.) or relocating meetings to more efficient smaller rooms. An interesting point is that irritation levels are examined in the study (irritation might be occur when an occupant gets a message by her Proxy agent frequently). According to examination, a simple message might reach much higher irritation levels than a complex message when the frequency of message sending is increased.

- Automation in Construction

Multi-agent comfort and energy system (MACES) provides an alternative model to manage and control building devices and occupants considering actual thermal zones, temperatures, occupant preferences, and occupant schedules. There are three types of agents which are device agents, human agents and meeting agents. Device agents represent HVAC, lighting and appliance agents. They mainly monitor and control sensors and switches. Human agents represent permanent and temporary occupants who possess different behaviours, schedules and preferences. The simulation of the system utilise four control strategies which are baseline, reactive, proactive and proactive-MDP. The control capabilities varies for each strategy from just adjusting temperature and lighting to relocating meetings. It is stated that proactive-MDP strategy in which agents additionally can change meeting schedules outperforms other strategies in terms of energy savings and comfort levels.

- A Multi-Sensor Based Occupancy Estimation Model for Supporting Demand Driven HVAC Operations

This paper focuses on the estimating the occupancy in buildings so as to adjust HVAC operations. The estimation model is built on several sensors. Each sensor node includes a light sensor, a sound sensor, a motion sensor, a humidity sensor, a CO₂ sensor, a temperature sensor and a PIR sensor. To estimate the number of occupants, sensor data is processed through radial basis function network. The overall detection rate is 87.62% for self-estimation and 64.83% for cross-estimation. The sensor node costs about $230 USD.

The following papers basically repeat the work explained in the above. However, they provide some brief descriptions of calculations utilised in the simulation.

- SAVES: A Sustainable Multi-Agent Application to Conserve Building Energy Considering Occupants

- Towards Optimal Planning for Distributed Coordination Under Uncertainty in Energy Domains

 

 

 

 

Heating Up

I read the following articles:

- A Negotiation Protocol for Multiple Interdependent Issues Negotiation Over Energy Exchange

This paper proposes a negotiation protocol in which off-grid households assumed to have renewable energy generation and electricity storage facilities exchange energy. Agents may benefit more through exchanging energy as they will reduce their storage losses. I liked the idea of the creation of a grid which is formed by off-grid households that generates their own energy and exchange the energy among themselves.

- Practical Distributed Coalition Formation via Heuristic Negotiation in Social Networks

The paper provides a novel framework for decentralised coalition formation in social networks. In the framework, agents can form coalitions without knowing peer's preferences through negotiations.  It is also demonstrated that the new negotiation strategy can increase social welfare by up to 10%.

- A Scoring Rule-Based Mechanism for Aggregate Demand Prediction in the Smart Grid

The mechanism called sum of others' plus(SOM) is developed to fairly distribute the savings obtained by agents. Unlike uniform mechanism which divides the savings equally among agents, SOM divides the savings according to each agent's contribution. This encourages agents to produce more accurate and costly estimates of future events.

- A panel model for predicting the diversity of internal temperatures from English dwellings

 

This work utilises panel methods to create a model which predicts internal temperature of households with more accuracy. The model considers both building stocks and human behaviour to predict internal temperature demand.  The number of occupants, household income, occupant age, building typology, building age, roof insulation thickness and wall U-value are some of the factors included in the model. The model can predict internal temperature of different building stocks within ~0.71°C at 95% confidence and explain 45% of the change of internal temperature among households.

 

- Use and usability of central heating controls

 

In this paper, the usability of heating controls is assessed through an experiment in which participants are asked to perform some tasks which represent the functionalities provided by the control. The result of the experiment states that heating controls are not well-designed enough to clearly express its all functionalities as the participants had difficulties to perform the most of the tasks.

 

 

 

Norman Thing

I finished reading the book "The Design of Everyday Things" by Norman, so I wanted to write some short notes to remind me later what I found interesting in the book. These notes do not cover everything mentioned in the book and I might miss some points since I only read it once. However, I can say that reading this book changed the way I see the world, which is strange. Although we live same things everyday, we generally cannot recognise them by ourselves unless we read or hear them from external sources. 

1- It is not our fault! The design has defect.

2- Conceptual models, feedback, constraints, affordance are the principles that a designer should focus on.

3- The Seven Stages of Action:

Forming Goal -> Forming the intention -> Specifying an action -> Executing the action ->
Perceiving the state of the world -> Interpreting the state of the world -> Evaluating the outcome

The greens are for execution while the yellows are for evaluation.

4-  What should be done for a good design?

• Utilise both knowledge in the world and in the head.
• Simplify the structure of tasks (balance depth and width).
• Make things visible: bridge the gap between execution and evaluation.
• Get the mappings right.
• Use the power of constraints (natural and artificial).
• Always consider errors and design for them.
• When all else fails, standardise (arbitrary).

Now, I started to read the book "Plans and Situated Actions" by Suchman.

Keep up with the Joneses

- Investigating Intelligibility for Uncertain Context-Aware Applications (Lim & Dey, 2011)

In this paper, the impact of intelligibility of context-aware applications is discussed. The intelligibility improves users' impresssions and their trust in the applications if the certainty of an application is higher enough and the action taken by the application is appropriate. Otherwise, when the uncertainty is higher, the intelligibility negatively affects users' impressions. This proposal is supported with an empirical evaluation in which the impressions of two mobile context-aware applications (LocateMe and HearMe) are investigated through applying different intelligibility and certainty levels.

The article above was about uncertainty and intelligibility of intelligent systems. From now on, I will post the documents that examines how social networks can be used to encourage energy efficiency, and I will start from Opower's online library. (http://opower.com/company/library)

- Energy efficiency through behavioural science and technology (Laskey & Kavazovic, 2011)

This paper is important to understand the success of Opower in the encouragement of home energy efficiency. Behavioral science, analytics and technology are utilised by Opower to provide effective customer engagement. In aspect of behavioral science, descriptive social-normative messages, which compare individualised energy use with the average use of neighbours, are used to influence individuals through the power of generated social norms. In order to prevent boomerang effect, complementary injuntive messages that approve good behaviours are also included in home energy reports. Moreover, the use of loss language is preferred as it is more powerful than gain language. Advanced metering technologies make the huge amount of comprehensive data available. Hence, the use of analytics is key factor to deal with the huge chunk of data. Opower exploits analytics to provide neighbour comparison, data disaggregation and personalised energy saving tips. From the technology perspective, scalibility and extensibility of the software play crucial role to meet new demands and to sustain users' trust. Moreover, Opower is conscious of the power of well-designed visual feedbacks having properties of visibility and affordance.

- Behaviour Change and Energy Use

The paper provides complementary insights to existing energy policies. It highlights how behavioural changes are important factors to reduce energy consumption and bills, and it states the three major benefits of behaviourally based changes. The first benefit is that the outcome is reached quickly unlike infrastructure changes which might take years. Secondly, behavioural changes are more cost-effective. Thirdly, citizens can directly benefit from the changes through savings. Hence, better understanding of how people behave in their everyday lives is essential. The paper represents three useful insights related to human behaviour.

1- Many people are keen to benefit immediately. Therefore, getting small rewards in a short term is more attractive than obtaining bigger rewards in a long period of time.

2- Social norms heavily effect individuals' behaviour. People learn how to behave through observing other people around them.

3- Individuals are tend to use default settings, so changing these defaults in a way to reduce consumption provides considerable energy saving.

 

 

First Step

So far, I attended induction programme and arranged my work environment including a new computer and a desk. I had first meeting with my supervisors. Roughly, my PhD will combine two research topics which are multi-agent systems and human-computer interaction. The aim of our research is to develop a well-designed platform in which multiple users and agents interact each other to overcome the peak demand problem of electricity energy market. The topics which are coalition formation algorithms, prediction techniques, game theory, interface designs and  usability tests will be examined throughtout the study. The research topic will become more precise after completing required literature reviews.

 

The papers that I read are as follows:

- Coalitional energy purchasing in the smart grid (Vinyals et al., 2012)

The paper introduces an algorithm to form efficient and stable coalitions which are composed of electricity consumers. The aim of coalition formation is to enable individual consumers to collaborate and benefit from bulk purchasing of electricity. By doing so, suppliers can better estimate load demand and adjust their production effectively, which results in low electricty costs for consumers.  The algorithm considers both spot market price and forward market price for coalition value calculation and takes account of social relations among consumers for coalition formation. The emprical evaluation of the coalition formation model demonstrates the following outcomes:
  . The density of social network, which is evaluated as the ratio between the number of relations and the number of consumers, affects the stability of coalitions. When the density increases, the stability of coalitions gets more fragile.
  . The difference between spot market price and forward market price is important factor to form larger coalitions. The future load demand can be estimated more precisely through the purchases of larger coalitions.

- On coalition formation with sparse synergies (Voice et al., 2012)

The article basically points out the computational redundancy in existing coalition formation algorithms. The redundancy occurs as the existing algorithms assume that all coalitions are feasible to be formed, although there might be constraints (sparse synergies) among coalition members. The article presents novel coalition enumaration and evaluation algorithms where computations are justly shared among agents to increase performance. Moreover, the article introduces an algorithm for coalitional structure generation, which is based on the enumaration and evaluation algorithms. The theoretical and emprical evaluation of the study shows that the algorithms do not include redundant computations and outperform the existing algorithms. 

- Competing with humans at fantasy football: team formation in large partially-observable domains (Matthews & Ramchurn, 2012)

I read this article since I am specially interested in playing football and the solution provided by the article might be useful in energy domain as well. The study generates an agent which acts as a football team manager in an online fantasy football game called Fantasy Premier League. The challenge is that the agent (manager) needs to sequentially form an optimal team for each game while there are constraints and uncertainties over players. The manager's decision problem is modeled as partially observable Markov decision process in which the agent needs to maintain the probability distribution over a set of possible states through observations. Here, the possible states correspond to the possible actions that the manager can perform, and the observations are the outcomes of played (or sampled) games and available statistics. The algorithm used in the model is based on model-based Bayesian reinforcement learning which permits more observable quantification of uncertainty and Q-learning that iteratively explores all action space to learn the best qualified actions. The solution is evaluated through created various managers who differ each other in terms of the consideration of the number of future games, the number of generated team samples and the depth of consideration of uncertainty. The outcome of the evaluation shows that Bayesian Q-learning outperforms other attempts in terms of final score.

 

- Toolkit to support intelligibility in context-aware applications (Lim & Dey, 2010)

In this paper, the necessity of intelligibility of context-aware systems is emphasised. The intelligibility is important for users to build trust to the systems. The paper reviews prevalent decision models like Rules, Decision tree, Bayesian Models and Hidden Markov Models. In addition, explanation types that users need to receive are stated. The paper represents a toolkit which automatically generates explanations for context-aware systems. The toolkit extends the Enactor framework through adding four components which are Explainer, Querier, Reducer and Presenter. In a nutshell, Explainer creates explanation structures, Querier indicates questions and confine explanations, Reducer filters complex explanations to simplify them, and Representer delivers the explanations to users or subsystems. The validation of the toolkit is demonstrated through utilising existing datasets from various context-aware systems in which each system uses different decision model. The study offers a promising approach to investigate how the generated explanations improve user understanding and trust for intelligent applications.

 

- Handling of uncertainty in interactive artificial intelligence systems (Hodgson)

This report focuses on the question "Is it possible to improve performance and usability of intelligent systems through interacting with users when there is uncertainty?". To provide an answer to this question, the report proposes two approaches. The first approach delegates decisions to users when there is a high level of uncertainty and agents cannot decide how to react properly. Another approach examined in the report ralates with the visualisation of uncertainty to users. Several display techniques of uncertainty are sampled in the report. The first approach is mostly proved to increase the effectiveness of intelligent systems. However, the effectiveness of the second approach is not certain. It might not be suitable for every domain such as aviation to display uncertainty since the display may complicate the situation.

 

I have also started to read the book "The Design of Everyday Things" by Norman.