{机器学习}【学习笔记】

keras库

1.安装方法：

• sudo/管理员身份运行cmd 
• pip install tensorflow 
• pip install keras 

---

论文学习A Deep-Reinforcement Learning Approach for Software-Defined Networking Routing Optimization

Realize: a DRL(deep reinforcement learning) agent 

Improvement:

• provide solutions for unseen network states(contrast to traditional table-based RL agents)
• provide a near-optimal solution in one single step

Feature(RL agent):

• off-policy
• actor-critic 
• deterministic policy gradient algorithm 

DRL agent:

• iteratively improving its knowledge of the relationship between 3 signals by means of 2 deep neural networks (actor and critic)

Signals in RL:

• state: Traffic Matrix(TM, being the bandwidth request between each sourcedestination pair)
• action: tuple of link-weights (that determine the paths for all source-destination node pairs)
• reward: based on the mean network delay

Objective of the agent: 

• determine the optimal behavior policy π mapping from the space of states S to\ the space of actions A (π : S → A) that maximizes the expected reward r ∈ R (minimizes the network delay).

Experimental methodology:

• network topology: scale-free,14 nodes and 21 full-duplex links, with uniform link capacities and average node degree of 3
• 1000 TMs:10 traffic intensity levels (TI,ranging from 12.5% to 125% of the total network capacity). For each TI  generate 100 TMs (of equal total traffic, using a gravity model)
• 100,000 randomly-generated routing configurations