Load-balancing problems like those faced in call center networks can be difficult to solve. 

Answering calls quickly is crucial to efficient operations, however, and long wait times can result in poor customer satisfaction. When large volumes of incoming calls come in, call centers need to route those calls in a timely manner and pick up fast. Not doing so results in dropped calls, balked calls, and unhappy customers. 

Using a modified version of AnyLogic’s Interconnected Call Centers model, we show how Pathmind reinforcement learning can outperform baseline heuristics to improve call center performance with more accurate and faster optimization.


Five interconnected call centers receive calls simultaneously. When a call comes in, a center must decide whether to accept the call or forward it to another location in the network. 

A randomly-initialized threshold between 20 and 25 minutes triggering a balked call is established. This feature represents how real-life callers would respond to long wait times. Put simply, it means that callers whose wait times exceed that threshold are likely to hang up. A call is considered “balked” when they do.

The problem to be solved is figuring out the best way to route calls so that wait times and the number of balked calls are minimized.

Why Pathmind Was Needed

Call center operations need to respond dynamically when unexpected surges in calls occur. Reinforcement learning outperforms baseline heuristics when those surges happen, and it reveals surprising strategies for greater efficiency even under normal operations.

Screenshot of the Interconnected Call Centers model


We compared the performance of the Pathmind policy to three baseline heuristics: 

    • no call transferring (calls are answered by the call center that receives them)
    • shortest queue (calls are automatically sent to the call center with the shortest line)
    • most efficient call center (calls are forwarded to the call center that can process them the fastest)

Shortest queue showed the best results of the three heuristics, and the Pathmind policy outperformed it to cut wait times by an additional 9.6%. The Pathmind policy also resulted in a 3.7% reduction in balked calls.

The performance of Pathmind AI versus baseline heuristics in the Interconnected Call Centers model is a great example of how simulation and reinforcement learning can solve more complex load balancing problems than other optimization methods. 

When a limited number of resources are available to perform tasks efficiently, reinforcement learning can offer better performance and new insights. Other load balancing challenges in areas such as logistics, manufacturing, and laboratory test processing can see improved results with AI.

Project Resources