Sustainability in intelligent building environments using weighted priority scheduling algorithm

2.1. Building automation system

Garg and Bansal [18] and Agarwal et al. [1] used occupancy sensors to increase energy saving of building automation systems. Similar to this work, several approaches have been proposed to develop Heating, Ventilation and Air Conditioning (HVAC) systems and efficient lighting [11,13,29,30,36]. Assessment and practical applications of these approaches can be found in [12,17,32,38,40].

In another work, a system was proposed by Furey et al. [16] to solve the weakness of real time location systems. This system uses artificial intelligence (AI) techniques for making sense of smart prediction of the future locations for the historical user location. Their proposed system was shown to be able to enhance real time location system in terms of accuracy, yield, cost, latency, and predictive ability. The authors concluded that their system could be useful in home and building automation system to contribute to an effective context prediction in replacement of movement sensor or conjunction [16].

None of the aforementioned work, in fact considers the conflict perception. However, intelligent environment should able to cope and manage any deadlocks or unpredictable behavior among heterogeneous systems in intelligent building. Conflict detection and resolution in an intelligent building and home is still a fairly new topic [45]. Several research studies in resource management and intelligent environments have discussed conflict detection and resolution. The conflict resolution of environments is a complex process and even it is a critical in multi-agent based intelligent environments [31]. Moreover, there is no single approach that can be applicable to solve all kinds of conflicts [24,51].

2.2. Agent-related conflict

Lee [31] proposed a model for multi-agent based using the concept of collaboration design to solve environmental conflicts among activities, users, and physical settings. In their model, an object agent is devised to modify the setting of environment through other agent with means of communication and collaboration. The authors tested a set of hypothetical cases to demonstrate the type of conflicts and their resolutions. The results presented in the work showed that their model can enable the environment as an organization of multiple agents to perceive the user activity more intelligently and efficiently [31].

Alshabi et al. [2] surveyed conflict resolution methods and agent cooperation models. In this work, the authors proposed a multi-agent architecture in which each agent chooses its balance between autonomy and cooperation independently. To solve the conflict among agents, they also suggested three negotiation strategies, including chaining, arbitrary leader election and cloning [2].

Jacak and Pröll [23] proposed a heuristic approach to manage conflicts in multi-agent systems. In their proposed approach, each agent is able to react and perceive to the environment to execute an action for negotiating with other agents. The negotiation and coordination of actions are primarily performed to attain the common global goal with the aim to resolve and avoid the conflict in an intelligent multi-agent system [23].

A context-aware multimedia presentation system was designed and proposed by Kung and Lin [28]. Their system provides a learning and prediction module with an adaptation mechanism to user’s needs based on context information. That is, whenever the contexts change, the system provides a mechanism for the best adaption control. In such system, a conflict is detected and handled on the basis of context’s priority and weight to achieve quality of service requirements [28]. Reinisch et al. [44] also proposed a comprehensive multi-agent system for an intelligent building to reduce the energy consumption without avoiding user comfort [44].

2.3. Intention and interest conflict

Tuttlies et al. [52] proposed a framework called COMITY, which uses a priority conflict detection and resolution. The resolution of conflicts in their approach is carried out through conflict avoidance, which results from responses of applications to different user interests. In COMITY, the conflict manager is based on a context model that is fed from the information available in database to detect and adapt the application at the run time [52].

In another work, Armac et al. [4] specified component as services. Each service contains actions that indicate the state of the transition on resources. In this work, a conflict is detected when a service tries to transit a resource from one state that is previously set by another service. Whereas, a conflict is resolved based on the rule-based mechanism and priority [4].

A dynamic conflict resolution of various context-aware applications in intelligent environments was proposed by Park et al. [39]. The authors integrate user’s preferences and intentions as the value that can be assigned to a context attribute by actions that demand from application and express user intentions. This is defined as a cost function measured by the distance between resolved value and user preference. The resolved value is further verified based on information that minimizes all the user’s cost that is involved in conflicts [39].

Silva et al. [48 ,49] proposed a methodology based on client-server architecture model to detect and solve conflicts. This model configures and selects the suitable conflict resolution algorithm. The decision model used is based on application’s demand for resource consumption and quality of services. The quality of service is the satisfaction of users on the subject of achieving resolution result [48,49].

2.4. Resource conflict

Retkowitz and Kulle [46] argued that the conflicts among users require to be resolved by users manually. In this regard, they proposed an approach that considers dependencies between services that recognized as binding. In this work, binding policies, handling service bindings, constrains, and access control in a smart home are defined based on priority group [46].

In a similar work, Huerta-Canepa and Lee [22] proposed a resource management schema based on ad-hoc interaction for smart spaces. Their work is built on the assumption that only two tasks can execute at each time. However, avoiding conflict in this work is done through device control and conflict resolution is performed based on the execution cost of a given task rather than current task, i.e. the utility and user priority at resource level [22].

2.5. Policy conflict

As mentioned in [26], policies and priority schema with rules among entities are the common conflict resolution techniques that establish acceptable system behaviors for a given situation.

Syukur et al. [50] presented a strategy to detect and avoid conflicts based on context change in order to control the application behaviors. Their proposed strategy depends on the system’s situation with respect to considered context, the number of users, and type of services. The authors argued that the conflict detection and resolution could improve the system performance by providing quick responses to the user’s demands [50].

Another policy conflict detection method was proposed by Jiao et al. [25]. In this work, the authors tried to improve the time of routine, which takes to search for every rule in a policy rule set. To this end, they started by formalizing policy rules, context and policy conflicts. Then, they proceeded with constructing a formal concept lattice based on policies which were clustered into associated formal concepts. The formalization was further used to help propose an algorithm to detect conflicts. The results presented in their work showed that the concept lattice could be as an effective support in policy attribute values and reduction of the number of policy rules [25].

Capra et al. [7] proposed a method, called CARISMA, to use dynamic policy conflict detection and resolution. The authors believed that the conflicts could not be solved statically at the application design time, where they must be resolved in the runtime. Capra et al. [7] defined two types of conflicts: (i) Intra-profile conflict that exists inside the application profile running on a given device. (ii) Inter-profile conflict that exists among the application profiles running on various devices. They used a technique that relies on sealed-bid auction among applications to achieve a fair conflict resolution method [7].

2.6. Authority conflict

Masoumzadeh et al. [34] investigated authorization conflicts in various situations. An authorization conflict happens, when two or more polices permit or deny an access simultaneously in a given situation. For instance, a policy allows a resident to switch on the radio while another policy imposes silence. The authors proposed a graph-based approach to enable precedence establishment among authorizations in a conflict situation. In their proposed approach, conflict detection is statically performed, while conflict resolution is done at runtime. Specifically, in the detection phase, a potential conflict graph is statically constructed to simulate an actual conflict situation. Based on the created graph, the approach provides the resolution [34].

2.7. Intelligent building-related conflict

Majumdar et al. [33] proposed a method, called a smart meeting scheduling algorithm. The algorithm was designed to automate the meeting scheduling with a specific focus on energy saving while adhering to capacity constraints and time conflicts. The authors used a number of scheduling algorithms from greedy to heuristic methods to reduce the energy use [33].

In another work, Resendes et al. [45] proposed a formal framework based on constrain solving to automatically detect and resolve conflict situations at home and building systems. The authors assumed that when users interact with Home and Building Automation Systems (HBAS), their preferences and intentions happen to be different, causing conflict situations [45].

Rashid et al. [43] proposed a data scheduling scheme, called SHAAL, to manage and process data generated from sensors to support the decision making in smart home and living systems. The authors utilized priority scheduling in SHAAL to facilitate their system in a real setting deployment. Similar to our work, priority scheduling determines which process should be run first, according to its priority given. Therefore, the process with high priority is served prior to low priority ones. If two elements are received with the same priority, then they are served based on their order in the queue [43].

At last, it has to be said that all the aforementioned works mainly focus on the higher layer (i.e. Application layer) which deals with user preferences and intentions. However, it is important to note that this work neglects conflicts and unpredictable behaviors arisen in lower layers (i.e. Sensors layer) that might broadcast or even leave behind undetected and unresolved conflicts, as an outcome of the lower layer’s constraints. Inspired by this gap in the literature, our proposed approach takes into account lower layer issues as part of its conflict detection and resolution process. However, the following sections elaborate the proposed approach in details.

5.1. Heterogeneous systems and priority level

Table 1 demonstrates the system priority which is assigned based on low priority value. For example, fire alarms and CCTV are allocated with the most noteworthy priority level because of their tendency to wellbeing discriminating concern in an intelligent building’s domain. Gadgets connected with fire alarm and digital surveillance systems are given extraordinary significance in a typical building setting. Furthermore, the heterogeneous systems with the same level of priorities are activated as requested in the queue. As show in Table 1, the priority of the system is determined based on low value priority level and high value weight of the system. For instance, digital surveillance and fire alarms are assigned with the highest priority level due to their nature being safety-critical concern in an intelligent building environment. Devices associated with digital surveillance and fire alarm systems like CCTV cameras are given a great importance in a typical building setting. Moreover, the systems with the same priority level, the one has a high weighted value is triggered. The priority policy of the systems is shown in Fig. 4.

As shown in Fig. 4, the proposed model manages the conflict occurrence among the heterogeneous systems in an intelligent building prior to a making decision. The model is designed based on a weighted priority scheduling. It is often convenient to group events into priority classes and use priority scheduling among the classes. As long as there are runnable events in a high priority class, the priority scheduling algorithm runs each event for one quantum that never bothers about lower-priority classes. In a case, the high priority class is empty, the algorithm then runs the class events of high weighted priority. If the higher classes are empty, the algorithm runs lower class of high weight. As a decision maker, we used a NB classifier to inter-operate among heterogeneous systems. In this paper, we compared NB model with two other machine learning methods, namely multilayer neural network and decision tree C4.5 to emphasize the efficiency of the NB model.

5.2. Naïve-Bayesian classifier

As considered in our framework, we applied NB classifier with considering conflict resolution as a decision-maker to attain efficient communication’s response as well as to automate decision making among heterogeneous systems in an intelligent building.

NB classifier is a classification algorithm based on Bayes rule. It can predict probabilities of class membership. For instance, it predicts the probability of a given sample belonging to a particular class. Bayesian classifier is based on Bayes’ theorem and posteriori hypothesis. Besides, NB classifiers have an important assumption that an attribute value of the class is independent of other attribute values. This Naïve assumption is named “class conditional independence” and it is made to reduce the computation involved. This is why that it has been labeled as “Naïve”.

OPEN IN VIEWER

Fig. 4.

Weighted Priority Scheduling.

In this research, NB classifier is modeled based on the rule repository to dynamically make a decision and achieve a unified integration of heterogeneous devices in an intelligent building environment. The Event–Condition–Action (ECA) format of the rule for system interoperability is presented as follows:

on event1 if condition1 do action1;

After creating the NB model, in the on-line stage, the events are timely received and scanned. When a new event (s) arrives, the NB model will predict which sub-system should be activated (more detailed of the rules’s format and transforming to truth tables, and modeling of NB classifier with the five systems without considering conflict aspect were explained and presented in [47]).

The Naïve Bayes theorem and the computation probability of P ( Y | X ) can be formulated in terms of probabilities P ( X | Y ) , P ( Y ) , and P ( X ) as: (4) P ( Y | X ) = P ( X | Y ) × P ( Y ) P ( X ) . In this theorem, X = x 1 , x 2 , … , x m is a sample, that contains set of m attributes that are all conditionally independent of one another. In Naïve Bayesian terms:

In our intelligent building environment, which consists of five systems, we have five attributes. The first four attributes are features and the last one is a class attribute. As mentioned above, X contains four attributes that are conditionally independent of one another, given class attribute, Y. To emphasize the interoperability (i.e. Communication response time) among heterogeneous systems, two more machine-learning techniques (i.e. Multilayer neural network and decision tree learning) are applied to be compared with NB classifier. These two techniques are modeled in a similar way to NB classifier. The overview and details of these techniques are explained in Section 5.3 and Section 5.4 respectively.

A multilayer perception is a feed forward neural network with one or more hidden layer. In this research, the networks consist of an input layer of source neurons, one hidden layer of computational neurons, and an output layer of computational neurons. The input layer accepts input signals from the intelligent building and redistributes these signals to all neurons in the hidden layer. The output layer accepts output signals from hidden layer and establish the output pattern of entire network [37].

A neuron in hidden layer detects the features, where the weights of the neurons represent the features hidden in input patters. These features are then used by the output layer in determining the output pattern. Figure 5 shows the feed-forward neural network. A feed-forward neural network with one hidden layer and enough neurons in the hidden layers can fit any finite input-output mapping problem. Feed-forward networks consist of a series of layers. The first layer has a connection from the network input. Each subsequent layer has a connection from the previous layer. The final layer produces the network’s output. The multilayer perceptron is presented in Algorithm 1.

OPEN IN VIEWER

Fig. 5.

Feed-forward Neural Network.

In our work, multilayer perceptron was implemented in .Net framework to model all five systems including CCTV, EM, FA, MA and PA. In this multilayer perceptron, the input layer consists of four inputs whereas the output layer has only one input.

OPEN IN VIEWER

Algorithm 1

Multilayer perceptron

As depicted in Algorithm 1, the first step is to initialize weights, bias value, and systems information. Later on, the back propagation neural network is activated by applying inputs and desired outputs. Besides, the actual outputs of the neuron in hidden layer are computed as follows: (5) V = ∑ i = 1 n x i × w i + θ , where n is the number of inputs, and θ is the bias value that applied to the neuron. Later on, the inputs are passed to the sigmoid activation function to compute the outputs: (6) Y sigmoid = 1 1 + e − V . In weight training, the weights are updated in the back-propagation network propagating backward the errors associated with output neurons (the error gradients for the neurons in output layer and in hidden layer are computed). The steps are repeated until the selected error criterion is satisfied.

Decision tree learning is one of the most widely used and practical methods for inductive inference. It is a method for approximating discrete-valued functions, which is robust to noisy data and is capable of learning disjunctive expressions. This method has been widely adopted in expert systems for captivating knowledge. It is mainly used for inductive methods to the given rate of property of an unidentified object to decide appropriate categorization according to decision tree rules. Decision trees relegate instances by traversing them from root node to a leaf node in a top-bottom manner. From the top root node of the decision tree, the method examines the attribute intended by this node, and then it moves down the tree branch according to the attribute assess in the given set. At the sub-tree level the whole process is reiterated.

In this research, among decision tree techniques (e.g. ID3 and C4.5), C4.5 decision tree was applied. One limitation of ID3 in comparison to C4.5 is that it is overly sensitive to features with large numbers of values. However, C4.5 uses a metric called “information gain”, which is defined by subtracting conditional entropy from the base entropy as shown in Eq. (7): (7) Gain ( P | X ) = E ( P ) − E ( P | X ) . This computation does not produce anything new. However, it allows measuring a gain ratio which is defined in Eq. (8): (8) GainRatio ( P | X ) = Gain ( P | X ) / E ( X ) , where E ( X ) is the entropy of the examples relative only to the attribute.

The C4.5 algorithm has an enhanced method of tree pruning that reduces misclassification errors due to noise or too many details in the training data set. Like ID3, the data are sorted at every node of the tree in order to determine the best splitting attribute. C4.5 uses gain ratio impurity method to evaluate the splitting attribute [42]. Anyanwu and Shiva [3] discussed that decision trees are built in C4.5 by using a set of training data or data sets as in ID3. At each node of the tree, C4.5 chooses one attribute of the data that most effectively splits its set of samples into subsets that are enriched in one class or the other. Its criterion is the normalized information gain (difference in entropy) that results from choosing an attribute for splitting the data. The attribute with the highest normalized information gain is chosen to make the decision. The C4.5 algorithm then recourses on the smaller sub lists [3].

In contrast to ID3, C4.5 has a better performance in terms of model build time as well as search time [20,27]. The pseudo code of C4.5 algorithm is illustrated in Algorithm 2.

OPEN IN VIEWER

Algorithm 2

C4.5 algorithm

As shown in Algorithm 2, C4.5 is implemented recursively. At the first stage, the termination criteria are checked to see whether the algorithm satisfies it or not. Later on, if it does not get terminated, then the information theoretic criteria for all attributes (CCTV, EM, FA, MA and PA) are computed according to the information theoretic criteria to choose the best attribute. Subsequently, a decision node based on the best attribute is created and the dataset (truth table) based on the newly created decision node is split. Afterwards, for all sub-dataset that were split, the C4.5 algorithm is called recursively to get a sub-tree. As such, the obtained tree is attached to the decision node to select the best attribute. These stages are repeated for all attributes until the termination criteria are satisfied. Finally, the tree is returned as a model in order to make a decision for new receiving events.

After creating the model, to achieve the efficient communication response and automated decision-making among heterogeneous systems in an intelligent building, conflict situation should be resolved. Section 5.5 presents the conflict resolution approach.

Events generated by heterogeneous systems are synchronized and updated via conflict resolution module in an intelligent building and are screened based on their predefined priority. The Conflict resolution module is implemented based on weighted scheduling mechanism. This weighted-scheduling mechanism is known as Machine Learning Weighted Priority Scheme (MLWPS). The Pseudo code of the MLWPS’s algorithm is shown in Algorithm 3. MLWPS performs priority weighted based on data scheduling for all the events generated in the intelligent building environment. Each system in an intelligent building has a different priority level. Each subsystem has a different weight to be triggered in an intelligent building through the SOAP protocol. For instance, as refer to Table 1, the five systems in an intelligent building are scheduled based on their priorities. The system with high priority is placed in a high class event. Furthermore, the sub-system with the same priority is located and sorted based on its weight. The events are run based on high weighted priority class. When events are received from systems via the SOAP protocol, the model concurrently checks the queue to see whether there exist events of high priority or not. If the received event is of the highest priority, then the model checks its associated weight to select the event with high weight. Otherwise, the event(s) will be queued based on priority and weight.

OPEN IN VIEWER

Algorithm 3

Heterogeneous systems Machine Learning Weighted Priority Scheme (MLWPS)

If the received event(s) is of lower priority compared to the queue event(s), then the received event(s) will be queued based on the priority and weight and consequently the event with high weighted priority from the queue will be pulled to trigger. It is worth mentioning that these generated events are jointly executed, when they match the priority level with event weights as shown in Table 1.

To verify our proposed model and its validation, experimental results and its analyses compared to the benchmark works [41] are presented in Section 6. Moreover, the results of NB classifier without considering conflict resolution has been published in [47].

6.1. Response-time analysis

To evaluate performance of the proposed model, five systems are measured together to express the capability of inter-operation among systems such as audio management system, fire alarm system, energy management system, main access door and digital surveillance systems, which are configured and tested via LAN and dedicated building server. The description of system devices is illustrated in Table 3.

For effective inter-operation of heterogeneous subsystems, a response time is significant and crucial metric for an intelligent building environment. A response time signifies the time to send and receive a message from the subsystem to the framework and back to subsystem [35]. This measurement will indicate the time required for marshalling the arguments and the return values of each subsystem’s message. Response time in an intelligent building environment points towards the maximum time of subsystems control tasks would take by execution without any interrupts or loads. The response time test performed to show the performance of the framework’s control mechanism in order to all systems talk to each other in a federated manner. Moreover, considering conflict resolution, which effects on response time, average waiting time of the events in the queue is calculated:

Therefore, W for each ready queue Q i is: (9) W Q = ∑ i = 1 q ( e Q i × P t i ) − P t i ∑ i = 1 q e Q i , 1 ⩽ q ⩽ 5 . If q = max ( ♯ of Q ) , the total average waiting time is calculated. Although what is calculated is amount of time an event(s) has been waiting in the ready queue, it is needed to compute the amount of time to trigger a particular event which is called turnaround time. Therefore, an average turnaround (end-to-end running time) is formulated as shown in Eq. (10): Average turnaround time ( TT ) = Average waiting time ( W ) + last triggered time ( P t ) therefore, (10) TT = ∑ i = 1 q ( e Q i × P t i ) ∑ i = 1 q e Q i , 1 ⩽ q ⩽ 5 .

As mentioned earlier, in the experiment, five systems were selected to evaluate the proposed framework in terms of its ability handling inter-operation among systems. The subject systems included one audio management system (public address), one fire alarm system, one energy management system, one main door sensor system, and one digital surveillance system (CCTV). The systems were configured and tested via LAN in the dedicated building server. The specification of system devices is described in Table 3.

For the performance evaluation, the mechanism of inter-operation is provided by the NB interoperability framework that is benchmarked using SOAP protocol’s response time on message exchange. The response time in an intelligent environment is an essential to make sure smooth functional of services with increasing growth of devices and sensors.

OPEN IN VIEWER

Fig. 6.

Average response time (ms) of three models with conflict resolution for six runs: (a) NB, (b) MLNN, (c) C4.5.

OPEN IN VIEWER

Fig. 7.

Average waiting time (ms) of three models for six runs: (a) NB, (b) MLNN, (c) C4.5.

In the experiment, we evaluated the approaches with considering conflict resolution. Figure 6 shows the response time of models with conflict run on five systems. The results were experimented based on three models, i.e. Naïve Bayesian classifier, multilayer neural network, and C4.5 decision tree for six runs. On the first run, the number of received events was 10. Consequently, the number of events increased to 20 for the next run. For the remaining four runs, the number of events increased by 10 for each consequent run. It is important to note that the Fig. 6 shows that by increasing the number of events the response time curve will climb linearly. The detailed results of the response time with conflict resolution between heterogeneous systems in intelligent building are given in Table 4.

In this research, as shown in Table 4, NB classifier has a comparable or better processing time ( Pt ), waiting time (W) and response time compared with the other two models, MLNN and C4.5. Among machine learning approaches, NB classifier, is a proper method to be used as a decision model where simplicity, robustness, and fast convergence compared to discriminative models (e.g. Logistic regression) is sought. Furthermore, NB classifier works well with real complex problems [53]. It also has to be said that the NB classifier works efficiently with minimum training data to assess the parameters (i.e. Mean and variance of attributes) with strong assumption, which is an added advantage over peer models. In NB classifier, when attributes are independent of each other, only the variance of the attributes for each class is required to be determined but not the entire covariance sample [9,14,19]. Therefore, it is made to reduce the computation involved.

Figure 7 and Fig. 8 show, respectively the average waiting time, and the response time of each system with conflict resolution for three machine learning models, i.e. NB, MLNN, and C4.5.

OPEN IN VIEWER

Fig. 8.

Average response time (ms) of three models with conflict resolution for each five Systems: (a) NB, (b) MLNN, (c) C4.5.

In contrast to the benchmark work (i.e. Event–Condition–Action (ECA) based model in [41]), our proposed model shows to have a better performance in terms of resolving conflict among heterogeneous systems in an intelligent building. Figure 9 depicts the results in a support of this out-performance.

OPEN IN VIEWER

Fig. 9.

Average response time (ms) of proposed models in comparison with ECA-Based model for the five Systems: (a) NB, (b) MLNN, (c) C4.5.

As shown in Fig. 9, the response time of five heterogeneous systems in an intelligent building is given based on their respective system types. The approximate response time value of five heterogeneous systems of Event–Condition–Action (ECA) based model and the proposed model with conflict resolution is presented in Table 5.

In comparison with the Event–Condition–Action based model (i.e. Work done in [41]), our approach improved the response time among the heterogeneous systems in an intelligent building even with conflict resolution. Although our approach declined the response time due to the conflict resolution process in comparison with the NB model that does not consider conflict resolution (i.e. [47]), we attained the efficient communication response as well as automated decision-making among heterogeneous systems in an intelligent building environment. In the ECA-based model and other existing rule-based models, when an event comes through the XML SOAP protocol, the model must consider and go through the entire rules to make a decision. This in turn could be a time-consuming and error-prone process. On the other hand, sending events to the central rule base makes it necessary to establish a connection to the database in the server as well as rule base center. However, such a connection cannot be guaranteed over a longer period of time. The central event detector could respond to these events with delay [21]. In contrast to the rule-based model, the proposed approach makes a decision based on a model that was trained once from the rules repository. This is to say that, the decision making support of the proposed approach in terms of response time is much faster than the previous work (compared to the ECA-based model in [8]). In addition, the proposed approach provides conflict resolution to support decision making which is meant to be more reliable and efficient among heterogeneous systems in an intelligent building. It is also important to note that, the results of the experiment showed that the systems with conflict resolution would still work efficiently.