BCG_MIN manual page
Table of Contents

Name

bcg_min - minimization or reduction of normal, probabilistic, or stochastic labeled transitions systems (LTS) encoded in the BCG format

Synopsis

bcg_min [bcg_options] [-strong | -branching | -divbranching | -observational | -sharp [-total | -partial | -gate] sharp_filename] | -divsharp [-total | -partial | -gate] sharp_filename] [-normal | -prob | -rate [-self]] [-epsilon eps] [-format format_string] [-class class_file] input.bcg [output.bcg]

where bcg_options is defined below (see GENERAL OPTIONS).

bcg_min takes as input the BCG graph input.bcg, minimizes this graph according to some bisimulation relation, and writes the resulting reduced graph to output.bcg, replacing input.bcg if output.bcg is omitted. Minimization means that distinct states of output.bcg are always non-bisimilar; As a consequence, output.bcg is the smallest graph (in number of states and transitions) bisimilar to input.bcg.

Description

bcg_min implements various algorithms to perform reduction of graphs encoded in the BCG format according to strong bisimulation, branching bisimulation, divbranching bisimulation, observational equivalence, sharp bisimulation, or divsharp bisimulation. A graph input or output by bcg_min can be:

General Options

The following bcg_options are currently supported: -version, -create, -update, -remove, -cc, -tmp, -uncompress, -compress, -register, -short, -medium, and -size. See the bcg manual page for a description of these options.

Particular Options

The following options are also supported:
-strong      
Perform LTS minimization according to strong bisimulation [Par81]. On (Discrete or Continuous Time) Markov Chains and on Markov Reward Models, this reduction agrees with lumpability of [KS76,KSK76,Buc94,Hil94] (see ANNEX 1, ANNEX 2, and BIBLIOGRAPHY below). Default option.

-branching
Perform LTS minimization according to branching bisimulation [GW96]. It is worth noticing that the notion of branching bisimulation is rather meaningless for probabilistic systems. Not a default option.

-divbranching
Perform LTS minimization according to divbranching (shorthand for divergence-preserving branching) bisimulation [GW96]. Divbranching bisimulation differs from branching bisimulation only in the way cycles of internal transitions (also called divergences) are treated. It is known that all states traversed by a cycle of internal transitions belong to the same branching equivalence class. While divergences are eliminated in the LTS obtained by reduction modulo ordinary branching bisimulation, a self-looping internal transition is kept in each such equivalence class in the LTS obtained by divbranching bisimulation. Unlike branching bisimulation, divbranching bisimulation preserves inevitability properties. Like branching bisimulation, it is worth noticing that the notion of divbranching bisimulation is rather meaningless for probabilistic systems. Not a default option.

-observational
Perform LTS minimization according to observational equivalence [Mil89]. It is worth noticing that observational equivalence is computationally more expensive than branching bisimulation, so that reduction may fail even for graphs containing only few thousands of states. To reduce the risk of failure, in a first step the input graph is automatically minimized according to branching bisimulation. This is sound because branching bisimulation is a graph relation stronger than observational equivalence. However, this optimisation is not applied if the -class option is set, so that bcg_min can print the equivalence classes relatively to the states of the input graph, instead of the states of the branching minimal intermediate graph produced in the first step. This option cannot be combined with neither -prob nor -rate options. Not a default option.

-sharp [ -total|-partial|-gate ] sharp_filename
Perform LTS minimization according to sharp bisimulation (the variant of the bisimulation defined in [LMM20] that does not preserve the divergences present in input.bcg), using the rules defined in sharp_filename to partition labels between strong and weak actions. See Section FORMAT FOR SHARP LABEL PARTITIONING below for details on the syntax of sharp_filename. The -total, -partial, and -gate options specify the "total matching", "partial matching", and "gate matching" semantics, respectively.

In "total matching" semantics, the regular expressions contained in sharp_filename denote full labels (i.e., gates possibly followed by experiment offers). A label matches if it matches some rule in sharp_filename entirely.

In "partial matching" semantics, the regular expressions contained in sharp_filename denote substrings of labels. A label matches if it contains a substring that matches some rule in sharp_filename.

At last, in "gate matching" semantics, the regular expressions contained in sharp_filename denote gates. A label matches if its first word (called gate) matches some rule in sharp_filename. In this case, regular expressions in sharp_filename should not contain characters forbidden in gate identifiers (e.g., " ", "(", or "!").

Option -total is the default.

This option cannot be combined with neither -prob nor -rate options. Not a default option.

-divsharp [ -total|-partial|-gate ] sharp_filename
Perform LTS minimization according to divsharp bisimulation (the variant of the bisimulation defined in [LMM20] that preserves the divergences present in input.bcg). See -sharp option above for the partitioning of labels between strong and weak actions. This option cannot be combined with neither -prob nor -rate options. Not a default option.

-normal
Consider input.bcg as a normal LTS. With this option, labels of the form "rate %f" or "label; rate %f" or "prob %p" or "label; prob %p" are processed as ordinary labels, without special meaning attached. Default option.

-prob    
Consider input.bcg as a probabilistic LTS. With this option, each label of the form "prob %p" or "label; prob %p" is recognized as denoting a probabilistic transition with probability %p. bcg_min will stop with an error message if, for some probabilistic transition, %p is out of ]0..1], or if the probabilistic transitions going out of the same state have a cumulated sum strictly greater than 1. With this option, labels of the form "rate %f" or "label; rate %f" are processed as ordinary labels. Not a default option.

-rate [ -self ]
Consider input.bcg as a stochastic LTS. With this option, each label of the form "rate %f" or "label; rate %f" is recognized as denoting a stochastic transition with rate %f. bcg_min will stop with an error message if, for some stochastic transition, %f is less or equal to 0. If the -branching or the -divbranching option is selected, and some state has both an outgoing stochastic transition and an outgoing internal (i.e., "tau") transition, bcg_min will print a warning and remove the stochastic transition in order to preserve the notion of maximal progress. With this option, labels of the form "prob %p" or "label; prob %p" are processed as ordinary labels. Not a default option.

If -self sub-option is given, all self loops (i.e., transitions that remain within the same equivalence class) having labels of the form "rate %f" are removed. This implements the weak Markovian equivalences described in [Bra02] and [BHKW05]. Not a default sub-option.

-epsilon eps
Set the precision of floating-point comparisons to eps, where eps is a real value. When eps is out of [0..1[, bcg_min reports an error. Default value for eps is 1E-6.

-format format_string
Use format_string to control the format of the floating-point numbers contained in transition labels (these numbers correspond to the occurrences of %f and %p mentioned in section DESCRIPTION above). The value of format_string should obey the same conventions as the format parameter of function sprintf(3C) for values of type double. Default value for format_string is "%g", meaning that floating-point numbers are printed with at most six digits after the "." (i.e., the radix character). Other values can be used, for instance "%.9g" to obtain nine digits instead of six, or by replacing the "%g" flag with other flags, namely "%e", "%E", "%f", "%G", possibly combined with additional flags (e.g., to specify precision).

-class class_file      
If class_file is the character '-', then display the equivalence classes on the standard output. Otherwise, display the equivalence classes in a file named class_file. Not a default option.

Note: In bcg_min versions up to 2.1, option -class was not followed by a class_file argument and equivalence classes were always displayed on the standard output. The class_file argument was introduced in bcg_min version 2.2. Because such evolution breaks backward compatibility, bcg_min issues an error message and stops if the argument following option -class is either an option (i.e., starts with a hyphen) or a file name with extension .bcg, which prevents overwriting an existing BCG file.

Note: Options -strong, -branching, -divbranching, -sharp, -divsharp, and -observational are mutually exclusive. If they occur simultaneously on the command line, the option occurring last is selected.

Note: Options -normal, -prob, and -rate (with or without -self sub-option) are mutually exclusive. If they occur simultaneously on the command line, the option occurring last is selected.

Unreachable States and Transitions

By principle and for efficiency reasons, bcg_min does not check whether states and transitions are reachable, neither in input.bcg nor in output.bcg.

As a consequence, output.bcg may contain unreachable states and transitions, meaning that input.bcg itself contained unreachable states and transitions. The converse is not true. In particular, if every state unreachable in input.bcg is equivalent to one that is reachable, then all states and transitions of output.bcg are reachable.

There is a single exception to this principle. Indeed, bcg_min may itself remove stochastic or probabilistic transitions, either due to maximal progress (when bcg_min is called with the -rate option) or because the value of a stochastic or probabilistic transition is less or equal to the precision of floating-point comparisons (when bcg_min is called with either -prob or -rate option, see the -epsilon option). In this case, bcg_min performs reachability analysis to eliminate those states and transitions made unreachable by such transition removal. If input.bcg contained states and transitions that were already unreachable before transition removal, then those states and transitions are also removed by the reachability analysis.

Note however that reachability analysis is not performed when bcg_min is called with the -class option, so that the unreachable states are listed explicitly in the class file. In that case, output.bcg may have more states and transitions than without the -class option, the additional states and transitions being unreachable.

If needed, unreachable states and transitions can be removed from input.bcg or from output.bcg by using the tools bcg_open and generator. See the bcg_open and generator manual pages.

Format for Sharp Label Partitioning

Sharp and divsharp bisimulations assume a partitioning of labels between so-called strong and weak actions [LMM20]. This partitioning is made through a file, which must respect the following grammar:
<sharp-set>        ::= <positive-header> '\n' <label-list>
                   |   <negative-header> '\n' <label-list>
<label-list>      ::=  <label>
                   |   <label> '\n' <label-list>
<label>           ::=  <regexp-denoting-a-label>
<positive-header> ::=  'hold'
<negative-header> ::=  'hold all but'

If the header is the positive header (hold), then any label matching a regular expression of the file will be considered to be a strong action. On the contrary, if the header is the negative header (hold all but), then any label matching a regular expression of the file will be considered to be a weak action. See the regexp manual page for details about regular expressions.

There is no mandatory suffix (i.e., file extension) for this file: any file name can be used; however, it is recommended to use the suffix ``.hold''.

Environment Variables

See the bcg manual page for a description of the environment variables used by all the BCG application tools.

Exit Status

Exit status is 0 if everything is alright, 1 otherwise.

Authors

Version 1.* of bcg_min was developed by Damien Bergamini (INRIA/VASY), Moez Cherif (INRIA/VASY), Hubert Garavel (INRIA/VASY) and Holger Hermanns (University of Twente). Pepijn Crouzen added -self sub-option. Version 1.* of bcg_min used the following algorithms:

Version 2.* of bcg_min was developed by Frederic Lang (INRIA/VASY). It uses (sequential) variants of the signature-based algorithm of [BO03] to compute strong, branching, and divbranching bisimulation on normal, probabilistic and stochastic LTS.

Operands

input.bcg
BCG graph (input)

output.bcg
BCG graph (output)

input@1.o
dynamic library (input or output)

Files

$CADP/bin.`arch`/bcg_min
``bcg_min'' binary program

See the bcg manual page for a description of the other files.

See Also

bcg , sprintf(3C)

Additional information is available from the CADP Web page located at http://cadp.inria.fr

Directives for installation are given in files $CADP/INSTALLATION_*.

Recent changes and improvements to this software are reported and commented in file $CADP/HISTORY.

Bugs

Please report bugs to Hubert.Garavel@inria.fr.

Annex 1 - Probabilistic Models

The probabilistic LTS model used in bcg_min is general enough to support the following models, which can be considered as special cases of probabilistic LTS:
Discrete Time Markov Chains [Nor97]
The graph contains transitions of the form "prob %p" only.

Discrete Time Markov Reward Models [How71]
The graph contains transitions of the form "prob %p" to represent transitions not obtaining an impulse reward. State rewards are associated to states by "normal" transitions with source and target states being identical. The label indicates the state reward such as "reward 5". Impulse rewards are associated to probabilistic transitions by prefixing the "prob %p" label with a label indicating the reward obtained by taking this transition, as in "impulse 4; prob %p".

Alternating Probabilistic LTS [Han91]
The graph contains transitions of the form "prob %p", as well as normal transitions in such a way that there is no state possessing both normal as well as "prob %p" transitions.

Discrete Time Markov Decision Processes [Put94]
The graph contains transitions of the form "prob %p", as well as normal transitions in such a way that there is no state possessing both normal as well as "prob %p" transitions. Normal and probabilistic transitions strictly alternate, i.e. normal (resp. "prob %p") transitions are not directly followed by normal (resp. "prob %p") transitions. Uses an encoding of Discrete Time Markov Decision Processes into strictly Alternating Probabilistic LTS proposed in [Arg00].

Generative Probabilistic LTS [GSS95]
The graph contains transitions of the type "label; prob %p" only, and for each state the sum of "%p" values leaving a state is equal to (or smaller than) 1.

Reactive Probabilistic LTS [GSS95]
The graph contains transitions of the type "label; prob %p" only, and for each state and each "label" the sum of "%p" values leaving a state is equal to (or smaller than) 1.

Stratified Probabilistic LTS [GSS95]
The graph contains transitions of the form "prob %p", as well as normal transitions in such a way that there is no state possessing both normal as well as "prob %p" transitions. Normal transitions are not directly followed by normal transitions.

Annex 2 - Stochastic Models

The stochastic LTS model used in bcg_min is general enough to support the following models, which can be considered as special cases of stochastic LTS:
Continuous Time Markov Chains [Nor97]
The graph contains transitions of the form "rate %f" only.

Continuous Time Markov Reward Models [How71]
The graph contains transitions of the form "rate %f" to represent transitions not obtaining an impulse reward. State rewards are assigned to states by "normal" transitions with source and target states being identical. The label indicates the state reward such as "reward 5". Impulse rewards are assigned to probabilistic transitions by prefixing the "rate %f" label with a label indicating the reward obtained, as in "impulse 4; rate %f".

Continuous Time Markov Decision Processes [Put94]
The graph contains transitions of the form "rate %f", as well as normal transitions in such a way that there is no state possessing both normal as well as "rate %f" transitions. Normal and stochastic transitions strictly alternate, meaning that normal (resp. "rate %f") transitions are not directly followed by normal (resp. "rate %f") transitions. Inspired by an encoding proposed in [Arg00].

Interactive Markov Chains [Her98]
The graph contains transitions of the form "rate %f", as well as normal transitions.

Timed Processes for Performance (TIPP) Models [HHM98]
The graph contains transitions of the form "label; rate %f", as well as normal transitions.

Performance Evaluation Process Algebra (PEPA) Models [Hil94]
The graph contains transitions of the form "label; rate %f" only. Passsive transitions are represented by abuse of the "rate" keyword. The transition label of a passive transition is augmented by a distinguishing string to indicate that the rate value has to be interpreted as a weight, such as in "THIS IS A PASSIVE TRANSITION LABELLED label WITH WEIGHT; rate %f". The actual distinguishing string used for this purpose is of no importance for bcg_min, but it must not contain ";" and must not start with the keyword "rate".

Extended Markovian Process Algebra (EMPA) Models [BG98]
The graph contains transitions of the form "label; rate %f" only. Passive and immediate transitions are represented by abuse of the "rate" keyword. The transition label of a passive transition is augmented by a distinguishing string to indicate that the rate value has to be interpreted as a weight, such as in "THIS IS A PASSIVE TRANSITION LABELLED  label WITH PRIORITY p AND WEIGHT; rate %f". The transition label of an immediate transition is augmented in a similar way, as in "THIS IS AN IMMEDIATE TRANSITION LABELLED label WITH PRIORITY p AND WEIGHT; rate %f". The actual distinguishing strings used for these purposes are of no importance for bcg_min, but they must not contain ";" and must not start with the keyword "rate".

Bibliography

[Arg00] P. R. D'Argenio. On the relation among different probabilistic transition systems and probabilistic bisimulations. CTIT Tech Report, to appear 2000.

[BG98] M. Bernardo and R. Gorrieri. A Tutorial on EMPA: A Theory of Concurrent Processes with Nondeterminism, Priorities, Probabilities and Time. Theoretical Computer Science 202, pp. 1-54, 1998.

[BHKW05] C. Baier, H. Hermanns, J.P. Katoen, V. Wolf. Comparative branching-time semantics for Markov chains. Information and Computation, vol. 200(2), pp. 149-214, 2005.

[BO03] S. Blom, S. Orzan. Distributed State Space Minimization. Electronic Notes in Theoretical Computer Science, vol. 80, 2003.

[Bra02] M. Bravetti. Revisiting Interactive Markov Chains. Electronic Notes on Theoretical Computer Science, vol. 68(5), 2002.

[Buc94] P. Buchholz. Exact and ordinary lumpability in finite Markov chains. Journal of Applied Probability 31(1), 59-75 (1994).

[GH02] H. Garavel and H. Hermanns. On Combining Functional Verification and Performance Evaluation using CADP. Proc. 11th Int. Symp. of Formal Methods Europe FME'2002 (Copenhagen, Denmark), LNCS 2391, July 2002. Available from http://cadp.inria.fr/publications/Garavel-Hermanns-02.html

[GSS95] R. J. van Glabbeek, S. A. Smolka, and B. Steffen. Reactive, generative, and stratified models of probabilistic processes. Information and Computation 121, pp. 59-80, 1995.

[GW96] R.J. van Glabbeek and W.P. Weijland. Branching Time and Abstraction in Bisimulation Semantics. Journal of the ACM 43(3):555-600, 1996.

[GV90] J. F. Groote and F. Vaandrager. An efficient algorithm for branching bisimulation and stuttering equivalence. Proceedings of the 17th ICALP (Warwick), LNCS 443, pp. 626-638, 1990.

[Han91] H. A. Hansson. Time and Probability in Formal Design of Distributed Systems. PhD thesis, University of Uppsala, 1991.

[Her98] H. Hermanns. Interactive Markov Chains. Ph.D. Thesis, University of Erlangen-Nuernberg, Germany, 1998.

[HHM98] H. Hermanns, U. Herzog, and V. Mertsiotakis. Stochastic Process Algebras - Between LOTOS and Markov Chains. Computer Networks and ISDN Systems 30, pp. 901-924, 1998.

[Hil94] J. Hillston. A Compositional Approach to Performance Modelling. Cambridge University Press, 1996.

[How71] R. A. Howard. Dynamic Probabilistic Systems, Vol II: Semi-Markov and Decision Processes. Wiley, 1971.

[HS99] H. Hermanns and M. Siegle. Bisimulation algorithms for stochastic process algebras and their BDD-based implementation. Proceedings of the 5th ARTS, LNCS 1601, pp. 244-265, 1999.

[KS76] J. G. Kemeny and J. L. Snell. Finite Markov Chains. Springer, 1976.

[KSK76] J. G. Kemeny, J. L. Snell, and A. Knapp. Denumerable Markov Chains. Springer, 1976.

[KS90] P. C. Kanellakis and S. A. Smolka. CCS expressions, finite state processes, and three problems of equivalence. Information and Computation 86(1), pp. 43-68, May 1990.

[LMM20] F. Lang, R. Mateescu, and F. Mazzanti. Sharp Congruences Adequate with Temporal Logics Combining Weak and Strong Modalities. Proceedings of the 26th TACAS, LNCS 12079, pp. 57-76, 2020.

[Mil89] R. Milner. Communication and Concurrency. Prentice-Hall, 1989.

[Nor97] J. R. Norris. Markov Chains. Cambridge University Press, 1997.

[Par81] D. Park. Concurrency and Automata on Infinite Sequences. In Peter Deussen (Ed.), Theoretical Computer Science, Lecture Notes in Computer Science vol. 104, pp. 167-183. Springer Verlag, March 1981.

[Put94] M. L. Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley, New York, NY, 1994.


Table of Contents