Improving Multi-agent Coordination by Learning to Estimate Contention