The training of autonomous agents often requires expensive and unsafe\ntrial-and-error interactions with the environment. Nowadays several data sets\ncontaining recorded experiences of intelligent agents performing various tasks,\nspanning from the control of unmanned vehicles to human-robot interaction and\nmedical applications are accessible on the internet. With the intention of\nlimiting the costs of the learning procedure it is convenient to exploit the\ninformation that is already available rather than collecting new data.\nNevertheless, the incapability to augment the batch can lead the autonomous\nagents to develop far from optimal behaviours when the sampled experiences do\nnot allow for a good estimate of the true distribution of the environment.\nOffline learning is the area of machine learning concerned with efficiently\nobtaining an optimal policy with a batch of previously collected experiences\nwithout further interaction with the environment. In this paper we adumbrate\nthe ideas motivating the development of the state-of-the-art offline learning\nbaselines. The listed methods consist in the introduction of epistemic\nuncertainty dependent constraints during the classical resolution of a Markov\nDecision Process, with and without function approximators, that aims to\nalleviate the bad effects of the distributional mismatch between the available\nsamples and real world. We provide comments on the practical utility of the\ntheoretical bounds that justify the application of these algorithms and suggest\nthe utilization of Generative Adversarial Networks to estimate the\ndistributional shift that affects all of the proposed model-free and\nmodel-based approaches.\n