Figure 3.

Reinforcement learning from zero prior knowledge. (a) Evolution of the CDR with laser chaos signals (sampling interval: 10 and 50 ps), quasiperiodic signals (sampling interval: 50 ps), coloured noise (sampling interval: 50 ps), and uniformly and normally distributed pseudorandom numbers. The CDR exhibits prompt adaptation with laser chaos when the sampling interval is 50 ps. (b) The CDR is evaluated as a function of the sampling interval from 10 ps to 400 ps, where the maximum performance is obtained at 50 ps with laser chaos. CDR with chaotic lasers yields superior performances compared to quasiperiodic and coloured noise signals. (c) Autocorrelation of the experimentally observed laser chaos, the quasiperiodic signal trains, and numerically generated coloured noises. The chaotic signals exhibit the negative maximum when the time lag is 5 or −5, exactly coinciding with the fact that the optimal adaptation is realized at 50-ps (10 ps × 5) sampling intervals.