Fuelled by the proliferation of smartphones, wireless traffic has experienced huge growth, which will continue with the emergence of ultra-broadband 5G applications, and exacerbate the capacity strain in cellular networks. Deployment of pico access points, reducing cell sizes and allowing more efficient reuse of limited radio spectrum, provides a powerful approach to cope with traffic hot spots and bring capacity relief. This network densification makes cell planning more challenging though, and tends to result in more irregular cells with possibly overlapping coverage areas and greater variability in traffic loads. This raises a critical need for more intelligent cell selection algorithms, which not only take signal strength values into account, but also load conditions in order to harness the full potential of the pico-cells. In the present paper we analyse online cell selection algorithms that use a parsimonious set of load-driven control parameters to determine an optimal user association in a measurement-based manner, without requiring explicit knowledge of the system parameters. We exploit stochastic approximation techniques to establish the convergence of the control parameters to the optimal values. Extensive simulation experiments for scenarios with many pico access points confirm that the algorithms are quite effective in optimally balancing the traffic loads in hot spot areas, and further demonstrate that they substantially outperform conventional approaches in terms of service denials and low throughput percentiles. We consider several implementation options and evaluate the relative benefits and potential tradeoffs.