Before pairing the terms Deep Learning and Optimization of Network Traffic it is worthwhile to revisit what we know about learning in the context of routing traffic – or: what we know about routing algorithms and their inherent optimization component. The latter already indicates that finding an optimal route involves techniques that seek to incrementally improve some value until it is found (or does not change anymore). To this extent it appears that the problem of finding good routes had involved a form of learning[1]Learning in the sense of looking at, comparing and ranking data structures being presented and identifying an optimal solution (optimization, finding a recipe). However, not so much in the sense of … Continue reading.
We want to revisit the main concepts here. We will also revisit proposals such as Q-Routing[2]Boyan, J. and Littman, M. (1994). Packet Routing in Dynamically Changing Networks: A Reinforcement Learning Approach.. And we want to round it up with recent proposals such as the MIND architecture[3]Geng, Y. (2017). MIND: Machine Learning based Network Dynamics. Talk at the Open Networking Summit 2016.. We also want to use theory from the dynamic control systems community on network architecture tradeoffs[4]Matni, N., Tang, A. and Doyle, J. (2015). A Case Study in Network Architecture Tradeoffs. ACM Sigcomm Symposium on SDN Research (SOSR) 2015. to see what the theory can do for us when different kinds of learning are involved. Please stay tuned.
References
| ↑1 | Learning in the sense of looking at, comparing and ranking data structures being presented and identifying an optimal solution (optimization, finding a recipe). However, not so much in the sense of taking in new data (experience, training on a feedback signal from the environment, a reflection about how one has done) which is used to alter the function in order to do better. The latter requires changes to oneself (learning!) which has to take place in time. The former is kind of fixed (static) and in a narrow sense is stripped off the possibility of taking in feedback for adaptation after the optimum has been found, thus is not learning – i.e. transforming – in a strict sense which we will apply from here on now. |
|---|---|
| ↑2 | Boyan, J. and Littman, M. (1994). Packet Routing in Dynamically Changing Networks: A Reinforcement Learning Approach. |
| ↑3 | Geng, Y. (2017). MIND: Machine Learning based Network Dynamics. Talk at the Open Networking Summit 2016. |
| ↑4 | Matni, N., Tang, A. and Doyle, J. (2015). A Case Study in Network Architecture Tradeoffs. ACM Sigcomm Symposium on SDN Research (SOSR) 2015. |