Recurrent neural network package for problems of time-series prediction and generation Copyright (c) 2009-2011, Jun Namikawa <jnamika@gmail.com> License: ISC license This package implements a gradient-based learning algorithm for recurrent neural networks. The package supports (1) both fully connected and sparsely connected networks, (2) both discrete-time neural networks and continuous-time neural networks, (3) training examples of both symbolic data and floating point numbers, (4) multi-threading, and (5) analyzing some characteristics (ex: Lyapunov spectrum, Kullback-Leibler divergence). === Installation === First, type `./autogen.sh' in the current directory to create configure file. Next, type `./configure' and when it finishes, type `make'. This will create `rnn-learn', `rnn-generate' and other utility programs. Run them with the argument `-h' to show the usages of them. If you wish to install the programs, type `make install'. By default, this will install all the files in `/usr/local/bin' or `/usr/local/lib'. You can change the install path with the `--prefix' option of the configure script, for instance `--prefix=$HOME' (use `./configure --help' for other options). === Requirements === Building this package requires a C compiler supporting C99 and Autotools (GNU Autoconf, Automake and Libtool). In addition, utility scripts in the `src/python' directory require python version 2.7 or 3.2. Gnuplot is also needed to run `rnn-plot-log' script. === Example === Here is a sample session. > cd examples > echo "import gen_target gen_target.print_sin_curve(500, 20)" | python > target.txt > rnn-learn -e 5000 target.txt > rnn-generate -n 1000 rnn.dat