2.1 Data aggregation policy
In the history of data aggregation policies, there are a little number of them focused on different aspects to design the aggregation policy in order to achieve energy efficiency.
According to the OSI layer model [9], most of the data aggregation policies focus on the network and the above layer, while some are implemented between the MAC and network layer [12]. Our research includes the LEACH [1] clustering algorithm, which acts as a network layer routing protocol in implementation, and therefore the whole work is constructed upon the network layer. This feature makes the data aggregation policies compatible for this clustering topology since both are based on the network layer. The basic idea behind these aggregation policies is to reduce the number of transmission and concatenate measurements thus decreasing header overhead.
The periodic per hop policy [6] is that each node waits for a predefined interval of time. After the end of the period, sensor nodes transmit all the received and sensed data.
The cascading protocol builds a distribution tree which tells each node that how many hops it is away the sink. The time limitation of the node is determined by the hops away from sink, further the node, shorter the timeout.
Algorithms proposed in [4][7][8] utilize the Markov Decision Process(MDP) model to solve the aggregation problem. A Markov decision process contains states, actions, transition processes, rewards and discount factor. These algorithms define its all these 5 elements with the purpose of maximizing the reward. In [7], the aggregation policy problem is regarded as an optimal stopping problem, which is a subset of stochastic sequential decision problems. A generalized MDP model, semi-Markov decision process
model, is used in [7] to construct the states and its transition distribution. The reward of the model is the aggregation gain, which is application dependent, with exponential decay.
Because the original states are too large for computing, the control limit algorithm is derived by simplifying the possible states to a computable level. This method calculates a threshold of the buffer size. If the buffer size is beyond the threshold, flush the buffer and transmit. Besides the control limit algorithm, two learning methods are proposed for comparison on the aspect of performance and computing complexity. Viewing from another aspect, Arroyo-Valles et al. in [8] gives every measurement an importance as priority, and transmits the measurements based on their importance. Importance is a general evaluation of energy, information source and time constraint. It could result in different performance because of different importance giving policy concerning information from other nodes. Yang et al. encapsulates the data aggregation policy into a transmission manager in [4], and the discussions combine theoretic proof and implementation for realistic scenario. This proposal outstands with a new suggestion that the reward of a measurement should decay linearly rather than exponentially. The solution is derived based on backward induction, a common technique to solve the Markov decision process problems. Since the deployment of WSN sensors should be costly and time-consuming, it is reasonable to share the facility with many applications, resulting with different deadlines. The previous works give decay on reward while it discards the expired measurements in [4]. This brings about the routing hops that should be seriously considered for. Therefore, multi-hop routing scenario is supposed and an enhanced algorithm is introduced.
These algorithms secure data against expiring at the cost of spending more communication energy. The energy efficiency, however, has more space of improvement at some cost of expiration. Our proposed method aims to achieve further energy efficiency
than the existing algorithms.
2.2 Clustering algorithm
LEACH [1] is adopted as the clustering algorithm in our experiments. Although being proposed as clustering algorithm, LEACH can also be implemented as a routing protocol. Some of the other clustering algorithms such as HEED [2] and DEEC [3] are the derivatives of LEACH. LEACH has four phases to consist a round. The first phase is to select the cluster heads in this round. LEACH gives a formula of the threshold, and the nodes generate a random number and compare to the threshold. The cluster heads broadcast their advertisements with the same energy level and the non-cluster-head nodes listen. The second phase is that the non-cluster-head nodes choose the cluster head to join.
After the first phase, the non-cluster-head nodes receive the advertisements from cluster heads. They choose the strongest signal as its source is the closest. The nodes then transmit to the selected cluster head in order to notice the cluster head the member information. The third phase is creating the transmission schedule for every node. In the fourth phase, the nodes transmit their data toward the base station. After these four phases are done, the protocol restarts from the first phase. LEACH wants every node to become cluster head once to share the extra loading of transmission and reception in a big round.
If there is N nodes, and average M cluster heads per round, the big round is comprised of N/M rounds. The node which has become a cluster head should wait until the big round end and the node can become a cluster head again.
For the above algorithms, experiments are made upon simple node distribution, such as linear topology and grid topology. However, in realistic application, the complex topology is different from those in experiments. Suggested that the performance of the
data aggregation policy may be affected by topology, there is an emergent need to discuss aggregation policy on some sophisticated topology, for example, cluster topology such as LEACH.