Recently, Visible Light Communication systems have gained a lot of attention. In this work we propose a context-aware and adaptive Visible Light Communication (VLC) system, able to dynamically react to the environmental changes in order to keep a good communication quality. In particular, we focus on a frame synchronization technique which is implemented by appending a preamble (repetitive insertion of sequences), to the transmitted data. At the receiver, a clean copy of the appended message is correlated with the received symbol stream for frame alignment. The size N (number of bits) of the preamble impacts on the performance of the communication system. Indeed, a short dimension of the preamble is to be preferred to reduce the control overhead (i.e. it is not carrying data information) but it could be not sufficient to perform a good carrier recovery, especially in the case of noisy environmental conditions. Since different external environmental conditions need different values of preamble length, the system must be able to gather information about its environment at any given time and adapt its behaviors accordingly (context awareness). Based on these premises, we propose a dynamic computation of N as ideal size of preamble for carrier recovery by modeling it as a multi-arm bandit problem and apply Thompson sampling to select in a fast and efficient way the best value of N [3]. Specifically, an agent tries to achieve as much award as possible by playing the most rewarding arm among J arms (J in our case represents the possible choices of the size N, that could be potentially unlimited but not all the sizes are meaningful, so we consider a limited sub-set). Each arm rewards randomly upon being played according to an unknown distribution. Our goal is the minimization of the exploration to find the most rewarding arm. The learning approach has been implemented to the receiver side. This choice is motivated by the fact that in the receiving path all data needed to implement the algorithm are known. We assumed that after the receiver computes the ideal value of N, it communicates this value to the transmitter that will consequently adapt the next frame. The algorithm has been implemented on a couple of low cost VLC prototypes consisting in an Arduino board, a driving circuit and a led array in the transmitting stage, a photo-diode, a trans-impedance amplifier and a second Arduino board in the receiving path. Transmitted signal is generated through software and the received one is processed through a proper ”Virtual Instrument”, developed using the commercial software LabView. Experimental results have shown the impact played by a correct choice of the parameter N on the reduction of the recovered carrier frequency variance and Bit Error Ratio (BER) in different environmental conditions.