Table 1 Summary

[23]

Action is of low importance in some states

The advantages and disadvantages of state and action, are respectively analyzed

Lessened the range of Q-value

[24]

Overestimation

Decomposing the max operation

More stable training results

[25]

Changed samples in experience replay

Improved the experience buffer training policy

Improved the performance of DDQN

[4]

Optimized execution time and cost

A pheromone update rule is designed

Better global search ability

[6]

Improved resource utilization, processing cost, and transmission time

The task scheduling is performed in two phases

Reduced makespan for tasks

[9]

Optimized load balancing

Each firefly flies towards a firefly that looks brighter than itself

Reduced transmission cost of workflow

[13]

Met the QoS requirements of users

Learned from its experiences without prior knowledge

Improved user satisfaction

[14]

Delay-sensitive task scheduling

Designed a reward function to reduce the average timeout period of tasks

Improved the scheduling efficiency of server-side tasks

[15]

Model-free policy for continuous action

Combines DPG and DQN

continuous action space

[16]

Scalable parallel tasks

A fully connected layer and an output layer

Improved task scheduling performance

[20]

Reduced energy consumption

Used the task priority to calculate the critical resources of the task graph

Reduced energy consumption in the data center

[26]

Optimize multiple objectives

Trained two DRL-based agents as scheduling agents

Reduced the average job duration

[27]

The efficiency of resource management

Proposed a blacklist mechanism

Converged quickly