NOTE: The CASC ReadMe file is in DOC/Readme. Short Installation Instructions for the impatient ------------------------------------------------- This assumes that you have GNU tar, sh and gawk in your search path! Simple installation of the first-order version (will install executables): > tar -xzf E.tgz > cd E > ./configure --bindir=/path/to/EXECDIR > make > make install > /path/to/EXECDIR/eprover -h | more Simplest installation (in-place): > tar -xzf E.tgz > cd E > ./configure > make > cd PROVER > ./eprover -h | more Read the rest of this file and the fine (if incomplete) manual if anything fails. There should be a copy of the manual in DOC/eprover.pdf. To build the higher-order version, user > ./configure --enable-ho > make > cd PROVER > eprover-ho -h The recommended command for running E on the file problem.p is eprover --auto --proof-object problem.p If you want to try the usually stronger strategy scheduling mode, use eprover --auto-schedule --tptp3-in --proof-object problem.p Replace "eprover" by "eprover-ho" for the higher-order-enabled version. You can add a time limit of 300 seconds with the option --cpu-limit=300 and a memory limit of 2 GB with --memory-limit=2048 for all "automatic" modes. You can reduce output with -s (or --silent). As of version 1.9.1, E will try to auto-detect the input format and adjust the output format accordingly. You can still force input and output formats via commandline options. The Equational Theorem Prover E =============================== This is the README file for version 2.3 "Gielle" of the E equational theorem prover. This version of E is free software, see the file COPYING for details about the license and the fact that THERE IS NO WARRANTY! This version of E is pre-released on an experimental basis. It supports a fragment of higher-order logic called lambda-free higher-order logic (LFHOL). Any issues you encounter with LFHOL support should be reported by email to petar.vukmirovic2@gmail.com. What is E? ---------- E is an equational theorem prover. That means it is a program that you can stuff a mathematical specification (in many-sorted first-order logic with equality) and a hypothesisconjecture into, and which will then run forever, using up all of your machines resources. Very occasionally it will find a proof for the conjecture and tell you so ;-). E has been created and is currently maintained by Stephan Schulz, , now with the help of several contributors (see DOC/CONTRIBUTORS). It is developed and distributed under the GNU General Public License. The E homepage can be found at http://www.eprover.org Installation: ------------- E can be installed anywhere in the file system, either by a normal user or by the system administrator. By default, the prover will still be compiled as a version that supports first-order logic only. To install the package, unpack the distribution (if you are reading this, you probably already did): gunzip -c E.tgz|tar -xvf - or (g)tar -xzf E.tgz (if you have GNU tar) This should create a directory named E. After unpacking, optionally edit E/Makefile.vars to your liking. In particular, if building for HPUX, comment out the suitable CFLAGS definition (for most systems, the default definition should be ok). Then change to the E directory: cd E Determine if you want to run E from it's own build directory or wether you want to install the executables in some other directory EXECDIR. In the first case, run ./configure otherwise ./configure --bindir=EXECDIR or, if you also want to install the man-pages into MANDIR, ./configure --bindir=EXECDIR --man-prefix=MANDIR To enable higher-order-support, add the option --enable-ho, e,g, ./configure --enable-ho Then type make to compile the libraries and all included programs under the E directory. If you want to install E in a particular EXECDIR, type make install You must have write permission in the EXECDIR, so if you install E outside your own home directory, you may need to become root or use sudo. Type make documentation to translate the LaTeX documentation (this requires LaTeX2e, pdflatex, and the packages theorem, amssymb and epsfig, which are included in most current LaTeX distributions). The manual should also be included as a pre-compiled PDF. For some operating systems, especially if you do not have the GNU gcc compiler installed, you may need to edit Makefile.vars manually to select tools and options. If you have any problems, look into E/DOC/PORTING. If you get into trouble, make rebuild will rebuild E completely to your current configuration. After installation, go to E/PROVER and type ./eprover BOO001-1+rm_eq_rstfp.lop to see the prover in action. Type ./eprover LUSK6.lop for a harder example. "./eprover -h" will give you some information and a list of options. For impatient people who do not want to read anything: eprover --auto --memory-limit=<80%_of_your_main_memory> should give a reasonable performance on a large class of problems (unless your main memory is really small). The auto mode will perform a heuristic pruning of the axiom set which may result in incompleteness for very large problems (many thousands of axioms). If you need completeness, use eprover --satauto --memory-limit=<80%_of_your_main_memory> In general, different proof problems are easy for different strategies. If you run eprover --auto-schedule --memory-limit=<80%_of_your_main_memory> or eprover --satauto-schedule --memory-limit=<80%_of_your_main_memory> the prover will try a series of strategies on the problem. It assumes a 300 second run time - if you impose a different one externally, it is important to let E know via the --cpu-limit=XXX option so that it can adjust the schedule. One of the features of E is the ability to produce semi-readable proofs. To use this, type eprover --proof-object By default, E will now automatically detect the input format (LOP, TPTP-2 or TPTP-3), and will select the matching output format (PCL2 for LOP and TPTP-2 inputs, TPTP-3 for TPTP-3 inputs). You can check the proof objects in PCL format for correctness using the tool checkproof in the same directory. "checkproof -h" should give you all necessary information. Note that checkproof cannot yet deal with the full first order part, and will skip anything not clausal. Also, support for independent provers in checkproof can be subject to bit-rot, as other systems and interfaces change. Bug reports and questions ------------------------- We welcome bug reports and even reasonable questions. If the prover behaves in an unexpected way, please include the following information: - What did you observe? - What did you expect? - The output of "eprover --version" - The full commandline that lead to the unexpected behaviour - The input file(s) that lead to the unexpected behaviour Most bug reports should be send to schulz@eprover.org. Bugs with the LFHO-version should be send to petar.vukmirovic2@gmail.com. Please remember that this is an unpaid volunteer service ;-).