Perfect channel estimation is very hard, time/ power consuming, and expensive; so it is not preferred (e.g. in mobile) communication systems. This paper seeks for new, cheap, low complexity, deep learning based solution. Several new combinations of deep learning and conventional structures (in different parts such as constellation shaper, channel estimator, and detector) are presented investigated, and compared over all atmospheric turbulence regimes from weak to strong (considering Gamma-Gamma atmospheric turbulence model). Results indicate that deep learning could provide close enough performance to the perfect channel estimation scheme, and it is immune to the atmospheric turbulence variation. The proposed deep learning based solutions are low cost, low complexity, with favorable performance. Accordingly, they are recommended for channel estimation in mobile communication systems. Because these system should deliver favorable, and cheap services to the costumers, which use a small mobile as transceiver that needs to be cheap, low complexity and low power consuming.Free space optical (FSO) communication system, due to its advantages over conventional radio frequency systems, is one the promising technology for far future communication services [1]. FSO is very effective in outdoor communication links, because it uses lasers as the transmitter and accordingly could support higher range communications properly without the need for amplification or correction [2]. Despite the fact that in outdoor communication, eavesdropping is easier, FSO link is immunized to it [3]. However, aside many advantages, the atmospheric turbulence of the outdoor environment significantly degrades performance of FSO system, and limits its practical applications [4]. Accordingly, channel estimation could really help improving performance of FSO system, and making it reliable. However, for this purpose, it is required to transmit a pilot sequence and use a processing complexity.Considering the complexity of the conventional communication systems, recently a new field of study emerged in optical communications, which by use of machine learning based algorithms tries to fulfil the cavity in complexity reduction of the existing techniques. Actually machine learning tries to remove the barriers by using the data itself. In the training phase, machine learns the structure of the data and finds the relation between input and output, then in the testing phase, the machine has the relation, so it could produce the desired output based on the input. The better training, the closer output to the desired. However, when the relation between input and output is very complex or when the input is not enough, it is required to learn the machine deeper and this is the idea of deep learning. Deep learning by consuming more complexity, could solve complex problems favorable.Recently, many investigations considered DNN for fiber OC applications such combating the Fiber effects [5], Modulation Format Identification [6], Optical Performance Monitoring [7], Opt...