A hierarchical fuzzy–genetic multi-agent architecture for intelligent buildings online learning, adaptation and control

Abstract

In this paper, we describe a new application domain for intelligent autonomous systems––intelligent buildings (IB). In doing so we present a novel approach to the implementation of IB agents based on a hierarchical fuzzy genetic multi-embedded-agent architecture comprising a low-level behaviour based reactive layer whose outputs are co-ordinated in a fuzzy way according to deliberative plans. The fuzzy rules related to the room resident comfort are learnt and adapted online using our patented fuzzy–genetic techniques (British patent 99-10539.7). The learnt rule base is updated and adapted via an iterative machine-user dialogue. This learning starts from the best stored rule set in the agent memory (Experience Bank) thereby decreasing the learning time and creating an intelligent agent with memory. We discuss the role of learning in building control systems, and we explain the importance of acquiring information from sensors, rather than relying on pre-programmed models, to determine user needs. We describe how our architecture, consisting of distributed embedded agents, utilises sensory information to learn to perform tasks related to user comfort, energy conservation, and safety. We show how these agents, employing a behaviour-based approach derived from robotics research, are able to continuously learn and adapt to individuals within a building, whilst always providing a fast, safe response to any situation. In addition we show that our system learns similar rules to other offline supervised methods but that our system has the additional capability to rapidly learn and optimise the learnt rule base. Applications of this system include personal support (e.g. increasing independence and quality of life for older people), energy efficiency in commercial buildings or living-area control systems for space vehicles and planetary habitation modules.

Introduction

The building industry uses the term intelligent, to describe the way the design, construction and management of a building can ensure that the building is flexible and adaptable, and therefore profitable, over its full life span. A definition which finds favour with many building managers and architects is that “An intelligent-building is one that provides a productive cost-effective environment through the optimisation of four basic elements; systems, structures, services, management and the inter-relationship between them” [20].

Computer scientists, however, hold a different view of intelligence. They are more concerned with giving machines management, analytic and control capabilities that are comparable to intelligent human activity. We prefer the definition––“An intelligent-building is one that utilises computer technology to autonomously govern and adapt the building environment so as to optimise user comfort, energy-consumption, safety and work efficiency”. In the context of a building, a system works by taking inputs from building sensors (light, temperature, passive infra-red, etc.), and using this and other information to control effectors (heaters, lights, electronically operated windows, etc.). If this system is to be intelligent, an essential feature must be its ability to learn and adapt appropriately. For this a system which can adapt and generate its own rules (rather than being restricted to simple automation) is required. According to Kasabov [16] an intelligent system should be able to learn quickly from large amounts of data therefore using fast training. An intelligent system should also adapt in real-time and in an on-line mode where new data are accommodated as it arrives. Also the system should be able to accommodate in an incremental way any rules that will become known about the problem. It should be memory-based, plus possess data and exemplar storage and retrieval capacities. It also should be able to learn and improve through active interaction with the user and the environment. It should have parameters to represent short and long term memory, age, forgetting, etc. It should also analyse itself in terms of behaviour, error and success. To the authors knowledge no system in the field of intelligent building has satisfied these conditions.

We view intelligent buildings as computer-based systems, akin to robots, gathering information from a variety of sensors, and using embedded intelligent agent techniques to determine appropriate control actions. In controlling such systems one is faced with the imprecision of sensors, lack of adequate models of many of the processes and the non-deterministic aspects of human behaviour. Such problems are well known and there have been various attempts to address them. The most significant of these approaches has been the pioneering work on behaviour-based systems from researchers such as Brooks [4] and Steels [23] who have had considerable success in the field of mobile robots. It might not seem obvious that a building can be looked upon as a machine; indeed “a robot that we live within”, but, in other work [6] we have justified the view that intelligent buildings, are analogous to robots, gathering information from a variety of sensors, and able to use behaviour-based techniques to determine appropriate control actions.

Fuzzy logic offers a framework for representing imprecise, uncertain knowledge. Similar to the way, in which people make their decisions, fuzzy systems use a mode of approximate reasoning, which allows it to deal with vague and incomplete information. In addition fuzzy controllers exhibit robustness with regard to noise and variations of system parameters. However, fuzzy systems have a well-known problem concerning the determination of their parameters. In most fuzzy systems, the fuzzy rules are determined and tuned through “trial and error” by developers, taking much iteration to determine and tune them. As the number of input variables increases (which is the case of intelligent buildings) the numbers of rules increase further magnifying the difficulty.

Evolutionary algorithms constitute a class of search and optimisation methods guided by the principles of natural evolution. Genetic algorithms (GA) are optimisation methods inspired by principles of natural evolution and genetics. GA has been successfully applied to solve a variety of difficult theoretical and practical problems by imitating the underlying processes of evolution such as selection, recombination and mutation. Their capability of learning enables a GA to adapt a system to deal with any task [12]. There are numerous reports in the scientific literature on designing fuzzy controllers using GA [2], [10], [15], [17]. However, most work uses simulation to overcome the time overhead problem caused by the large number of iterations needed for a conventional GA to develop a good solution. Thus it is not feasible for a simple GA to learn online and adapt in real-time. The situation is worsened by the fact that most evolutionary computation methods developed so far assume that the solution space is fixed (i.e. the evolution takes place within a pre-defined problem space and not in a dynamically changing and open one), thus preventing them from being used in real-time applications [16].

Our work is concerned with utilising an intelligent embedded-agent approach based on a double hierarchical fuzzy–genetic system, similar to the approach already taken in our previous work in mobile robotics [13], [14], to create an integrated and semi-autonomous building control system. There are other research projects concerned with applying AI to buildings most notably [5], [7], [8], [19]. Examples of such work include research in Sweden [8] that utilises multi-agent principles to control an intelligent building. Their primary goal is energy-efficiency, and although their system does adjust the heating and light level to suit individual preferences, these settings must be pre-defined. Their agents are built from traditional AI (i.e. not behaviour based) and their work does not address issues, such as occupant based learning. The system, so far implemented in simulation only, managed to achieve energy savings of 40% over the same building being controlled manually by occupants. A group in Colorado [19] are using a soft computing approach––neural networks––focusing solely on the intelligent control of lighting within a building, by anticipating when particular zones (regions in a room) will be occupied or unoccupied. Their system, implemented in a building with a real occupant, also achieved a significant energy reduction, although this was sometimes at the expense of the occupant’s comfort. They use a centralised control architecture which differs from our multi-agent approach and which includes no adaptation, as is the case for us. A third group based at the MIT Artificial Intelligence Lab in Massachusetts is working on an intelligent room project. They employ a mix of cameras, microphones and multiplexer to enable people to interface with room-based systems in a natural way using speech, gesture, movements, and context information [5]. This primary focus on facility of the user interface differs to our work where ideally the agent remains more or less invisible to the user of the building. There have been also other research projects by [1] and [11] concerned with producing optimal models for buildings which will used later in control, however these techniques lacks flexibility as if the system characteristics changes the system has to repeat a time consuming offline learning cycle. Also these techniques ignores the fact that the building control is subject to human desires which can vary from a person to another, so there is a need for online adaptive and interacting learning. We believe our approach is unique in seeking to apply a hierarchical fuzzy control architecture with genetic learning. This approach allows the learning or adaptation to be performed through interaction with the environment (with the occupant of the room being seen as part of the environment). In this approach there is no need for simulation or direct human intervention into the rule setting system thus satisfying the definitions of intelligent autonomous agents in robotics [23] and intelligent systems as explained above.

Broadly speaking, the agent work described here is situated in the recent line of research that concentrates on the realisation of embedded real-time agents grounded in the physical world. Traditional approaches such as real-time control and expert systems rely on predictive models and thus have difficulty in environments that are of intractable complexity (in terms of the dynamics and number of variables) being essentially (if not actually) non-deterministic. However, a number of robotic based researchers have had notable success in dealing with this type of problem based on techniques that can broadly be categorised by the term behaviour based systems. We take this approach as our starting point and have developed a method of implementing this as a hierarchical fuzzy system with numerous consequent advantages. Thus, our macro-control architecture belongs the “behaviour based control architecture” school pioneered by Rodney Brooks of MIT in the late 80’s [4]. In this approach a number of concurrent behaviours (mechanisms to attain goal or maintain state) are active (sensing environment, effecting machine) to a degree determined the relationship of the machine and environment. We have extended this approach by developing a number of mechanisms that address both behaviour-integration, in the form of a hierarchical fuzzy controller, and genetic learning, combined into a single architecture which we label an “associative experience engine” [13], [14]. In addition to dealing with control problems that are difficult to model, the self-programming nature of the system has benefits in reducing software development costs throughout the lifetime of the product. In this paper we give an overview of both the hierarchical fuzzy control and the genetically based associative experience engine applied to develop an embedded-agent within a distributed intelligent buildings architecture.