Open prashantsrivastava-131983 opened 5 years ago
RL optimizes the objective you specify (through the reward signal). If trading off smoothness for other factors (e.g., lower rebuffering rate) increases the total QoE, Pensieve will automatically do that. Increasing the smoothness penalty can help reduce the chance of bitrate switching but you might trade off other factors if the network bandwidth is limited.
Hi, I have trained the model on Ubuntu 16.04 with around 200000 iterations with 16 agents. Used default hyper params. Although, total QoE is higher than legacy buffer based approach, but there is smoothness issues in penseive. With every few seconds of test playback , it switches higher or lower. Is there a way to minimize switching. Would increasing smoothness penalty help in this case? Additional information: Testing Segment sizes - 2sec Threshold buffer duration - 25seconds
Any input in this regard would highly be appreciated.