Policy Learning for Adaptive Allocation of Cloud Resources to Multi-tier Web Applications

 Two experiments executed in three phases:  Exploration: we initialize an empty policy and let the system learn in real time  Exploitation: the agent simply uses the previously-learned policy to automatically resolve bottlenecks  Baseline: scale up every replicable tier every time a bottleneck occurs The workload pattern modeling method learns a clustering model: Update Q-value for each action/state pair The reward function:  encourages the agent on success  discourages the agent on failure to maintain the SLA  discourages overprovisioning Motivation