Of course most of you know the major disadvantages of gradient ascent/descent approaches. Most popular candidates are local optima and plateau problems.
In this post, I want to give you a simple example, that illustrates the plateau problem not just leading to slow convergence, but a complete fail of the search process (even with Rprop).
The next tournament on our schedule is the RoboCup World Championship! Japan was the original host country, but due to some real problems (I guess Fokushima and other Natural Disasters) they canceled it. Lucky for us is that the Netherlands spontanious step in, so that the next World Championship is only 3-4 hours by car away.
The official homepage is already online.
See you in Eindhoven on the RoboCup World Championship.