This paper discusses load balancing heuristics in a general-purpose distributed computer system. To minimize the mean response time of a task, every new task is scheduled to be executed either locally or at a remote host, depending upon the current load distniution. We implemented a task scheduler based on the concept of a stochastic learning automaton on a network of Unix workstations. The used heuristic and our implementation are shortly discussed. Creating an artificial, executable workload, a number of experiments were conducted to determine the effect of different workload descriptions. These workload descriptions characterize the load at one host and determine whether a newly created task is to be executed locally or remotely. Six one-dimensional workload descriptors have been examined. Also, two more complex workload descriptions were considered. The best results were obtained with a relatively simple workload description, the number of tasks in the run queue per host. Using more complex workload descriptions, in contrast, did not improve the mean response time, as compared to the best single workload descriptor.