This chapter covers a summary and conclusion of this dissertation as well as suggestions for future work. In this dissertation several well known machine learning techniques were applied to opportunistic DTN routing, including the reinforcement Q-Learning approach, an innovative multi-label classification bun- dle filter and a clustering based bundle filter for epidemic routing.
8.1 Summary
Chapter 1 discusses the introduction and motivations for cognitive and ma- chine learning based networking and how these relate to the thesis contribu- tions. Chapter 2 covered background material on the NASA SCaN Networks, the main concepts of DTNs and an introduction to the machine learning technique used in this work. Next, selected related works in DTN routing are discussed as well as other works that have applied related techniques such as Naive Bayes classifiers to DTN routing.
Chapter 4 discusses the approach taken to the contributions of this thesis, such as the development of a network and software architecture, and the pros and cons of distributed versus centralized learning architecture. From here, the
majority of the work in this thesis used a centralized architecture, other than Q- Routing which is a distributed approach. Each technique is then discussed in detail. The development of the classification approach including the determina- tion of suitable attributes, labels and problem formulation are discussed. Next, the development of the cluster based approach are covered.
In Chapter 5, the actual details of how these algorithms where implemented and tested are discussed. A variety of DTN bundle protocol implementations were investigated for the suitability of this work and IBR-DTN was selected as the best choice. The details of the modifications to IBR-DTN to integrate the machine learning models are covered. Chapter 6 discusses the development of a DTN testbed and the tradeoff between a simulation or emulation test envi- ronment. Several popular network simulators were explored and the final se- lection of the CORE emulator is discussed with details on how the emulation is performed using Linux containers.
Finally Chapter 7 covers metrics used to evaluate the learning and routing performance and a summary of results are given. It is shown that these route selection techniques are promising way to reduce overhead associated with epi- demic routing while still providing a comparable delivery ratio.
8.2 Conclusion
The techniques of clustering and classification are used to reduce the con- sumption of resources such as bandwidth, processing time and data storage. Us- ing historical data from the network, both techniques are able to predict which neighbors are the most likely to deliver bundles to their destination. For systems
with less data storage capacity, limited bandwidth or limited processing capabil- ities, the overhead associated with epidemic routing can be burdensome. Both machine learning techniques reduce the number of bundles that must be repli- cated, while still delivering a satisfactory number of bundles.
While Q-Routing is an interesting approach for many network scenarios, the long delays associated with some types of DTNs may make estimation of the re- ward function (end-to-end delay) and propagation of the estimate difficult. For this reason, alternative algorithms such as supervised learning (classification) where considered. It is still very promising to consider Q-routing and other re- inforcement learning techniques, however this will require additional consider- ation for DTNs. The use of reinforcement learning and Q-Routing with a cross- layer technique for obtaining information from lower protocol layers in particu- lar may be a promising approach.
8.3 Future Work
This section covers several interesting avenues that could be pursued for fu- ture work.
8.3.1 Cross Layer Approach
A cross layer approach as discussed in [93] and [32] could be used to incorpo- rate information from the lower protocol layers to add addition features to each reliability classifier for the nodes in the network. This information could also be used in a reinforcement learning scheme to reward routing decisions that pro- duce less errors or retransmission in the underlying protocol layers.
8.3.2 Evolutionary Clustering
Evolutionary clustering [94],[95] is a type of clustering which takes into ac- count points changing in time and updates clusters accordingly. This could po- tentially improve the performance of the cluster based filter which simply used a naive clustering method where all points within the period are clustered once and then matched back to the location and point in time.
8.3.3 Deterministic Routing
This work focused on an opportunistic DTN environment using replication routing techniques. Additional work could be done to apply machine learning techniques to networks that perform in a more deterministic manner with addi- tional constraints.
8.3.4 Radio/Optical Link Models
The emulations conducted used a very basic model of the wireless commu- nication links based on distance and data rate. However, EMANE has the ability to use custom defined radio models. The development of a higher fidelity radio model or development of an optical link model would be interesting. This was left for future work due to the scope of this dissertation being more focused on DTN software and routing algorithms.