Dynamic Sparse Training via Balancing the Exploration-Exploitation Trade-off