Learning Embeddings for Sequential Tasks Using Population of Agents