Overview
TAMER is a planner for the ANML (read as “animal”) planning specification language. The objective of TAMER is to provide a library of functionalities to model, solve and analyze planning problems in practice.
TAMER can be used in two ways. As a standalone, command-line planner to develop, debug and solve planning problems and as a library that can be embedded in applications offering planning functionalities.
ANML and TAMER
TAMER internal formalism and representation is adherent to the principles of the ANML (Action Notation Modeling Language) planning language. ANML is the default and more comprehensive textual input format for TAMER (but it is not the only one, also PDDL 2.1 is supported). See doc/ANML.md for a detailed description of the supported ANML fragment and extensions.
(Still) Unsupported Features
- HTN features
- Interval labels (e.g.
[all] contains g1: x;
) and interval constraints (e.g.g1 < g2
) - Set and vector types
- Uncertainty and set operators (except for
:in
in durations to express temporal uncertainty)
Note:
This document is still a draft and needs more detailed explanation of the language and its features.
Extensions
We added few syntactical extensions to the language.
Constants default values
ANML does not have a way to specify default values for function constants. For example, consider the following declaration.
constant integer c(float x);
ANML allows to initialize a finite number of points of constant c
. For example:
c(0.1) := 3;
c(3.14) := 2;
but there is no way to set a default value to the constant that is taken when a point where the constant has not been initialized is queried. We extended the ANML syntax to allow this specification.
c(*) := 0; \\ Indicates that c is 0 in all the points where the c function is not explicitly initialized
Moreover, for constant functions having multiple arguments, we allow for partial specification of the argument list to express default values. For example:
constant integer c(float x, integer y);
c(1.0, 2) := 4;
c(1.0, *) := 0; // Sets to 0 all the values of the function having 0.1 as first parameter (except for the points set explicitly such as (1.0, 2))
c(2.0, *) := 10;
c(*) := -1;
This feature is useful for representing the flattening of general ANML models. Consider the following example:
type T with {
constant integer c(integer x, float y);
....
}
instance T a, b;
a.c(*) := 9; // Pseudo-syntax for the default value
a.c(1, 3.14) := 0;
b.c(*) := 1;
b.c(1, 6.28) := 0;
If we want to express the very same model without using structured types we can produce the following model.
type T;
constant T_c(T object, integer x, float y);
instance T a, b;
T_c(a, *) := 9;
T_c(b, *) := 1;
T_c(a, 1, 3.14) := 0;
T_c(b, 1, 6.28) := 0;
This is the actual output of the Tamer flattener transformer that gets rid of the object-oriented syntactic sugar.
Note
The order of assignment instruction is not relevant: more specific assignments are always preferred to more general ones. Moreover, it is forbidden to assign multiple times the same value of the same constant.
Command-Line Usage
TAMER can be used as a command-line plannner. It can be executed from the system prompt as follows:
$> ./tamer (code-generate | convert-to | solve | transform | validate | trace-to-plan | repl) [options] <input-file-1> ... <input-file-n>
Tamer automatically combines all the given input files as a single problem. In this way it is possible to split the “domain part” and the “problem part” of the ANML instance in separate files and simply invoke:
$> ./tamer (code-generate | convert-to | solve | transform | validate) [options] domain.anml problem.anml
The order of the input files is not important.
Also, Tamer is able to read PDDL2.1 specifications using the --pddl
flag (see examples below).
Finally, if no input file is given to Tamer, the ANML input is expected on the command-line as standard input. This is useful to use the Tamer executable in a pipe with other tools (e.g. a script that outputs an ANML instance) as follows.
$> cat instance.anml | ./tamer (code-generate | convert-to | solve | transform | validate) [options]
Functions
- code-generate – generates the C++ code of a specialized planner and a specialized simulator for the input problem.
- convert-to – converts the input ANML problem in the specified target format.
- solve – solves the input planning problem.
- transform – transforms the input problem using any pipeline of the internal TAMER transformers.
- validate – validates the provided plan for the input problem.
- trace-to-plan – converts the input trace in a plan.
- repl – starts an interactive shell. With this function, the input files are Python files and they are executed before starting the shell.
Options
The options can be divided in the following groups.
Generic options
- logging level – to set the logging level. The level is a number between 0 and 5 with 0 being no-log and 5 being all the log. The default is 0.
[(--logging-level | -l) <level-value>]
- debug – to enable debug mode, letting exceptions to propagate to top-level.
[--debug | -D]
- pddl – to indicate that PDDL is expected as input specification language instead of ANML.
[--pddl | -P]
- transformer – to indicate the transformer to apply to the input problem. It is possible to specify multiple transformers and the order of specification will be the order of application.
[(--transformer | -t) (flattener, action-grounder, full-grounder, temporal-uncertainty-compiler, intermediate-effects-compiler, set-compiler, contains-compiler, partial-order-compiler, empty-conditions-compiler, usertype-fluents-compiler, bounder, constant-promoter, free-var-params-compiler, forall-compiler)]
- show problem – prints the planning problem after transforming it.
[--showproblem | -p]
- bounding problem – indicates the bounding problem used by the bounder transformer.
[(--bounding-problem | -b) <bounding-problem>]
- disable-prune-actions – Disable actions pruning in the grounding.
[--disable-prune-actions | -u]
code-generate options
- problem free – enables problem-free planner code generation. By defaul TAMER generates problem-specific code.
[--problem-free | -f]
- output planner – indicates the output file name where to save the generated planner code. The default is gplanner.cxx.
[--output-planner <output-planner-file>]
- output simulator – indicates the output file name to save the generated simulator code. The default is gsimulator.cxx.
[--output-simulator <output-simulator-file>]
convert-to options
- target – to indicate the target format for the conversion of the ANML problem.
[--target (ltl | smv | uppaal | pddl)]
- output – to indicate the output file name of the problem encoding in ltl or smv
[--output <output-file>]
- domain – to indicate the output file name of the PDDL domain.
[--domain <output-domain-file>]
- problem – to indicate the output file name of the PDDL problem.
[--problem <output-problem-file>]
- model – to indicate the output file name of the UPPAAL model.
[--model <output-model-file>]
- query – to indicate the output file name of the UPPAAL query.
[--query <output-query-file>]
solve options
- algorithm – indicates the name of the planning engine to be used. The options are ftp, rlftp, ctp, smt or tsimple. The default is ftp.
[(--algorithm | -a) (ftp | rlftp | ctp | smt | tsimple)]
- plan epsilon – to set the time quantum used to separate two actions in a partial-order-termporal plan. The default is 0.001.
[(--plan-epsilon | -e) <epsilon-value>]
- rational precision – to force the plan timing to have a rational precision.
[--rational precision | -R]
- time triggered – to force the produced plan to be a time-triggered plan.
[--time-triggered | -T]
- mission – to force the produced plan to be a mission plan.
[--mission | -M]
- smt max horizon – to set the maximal number of steps in SMT-based planning. The default is 100.
[(--smt-max-horizon | -H) <number-of-steps>]
- smt optimization rounds – to set the number of optimization steps in SMT-based planning. [Not implemented yet]
[(--smt-optimization-rounds | -O) <optimization-steps>]
- weight – to set the weight parameter for the tsimple and ftp planners. The value is a number between 0 and 1. The default is 0.5.
[(--weight | -w) <weight-value>]
- tsimple heuristic – to indicate the heuristic to be used in the tsimple planner. The options are hlandmarks or hplus. The default is hplus.
[(--tsimple-heuristic | -A) (hlandmarks | hplus)]
- tsimple goals serialization – to enable goals serialization.
[--tsimple-goals-serialization | -S]
- iw-disable-tuple-computation-in-initial-states – Disables the tuple computation in the initial states.
[--iw-disable-tuple-computation-in-initial-states | -c]
- ftp heuristic – to indicate the heuristic to be used in the ftp planner. The options are hadd, hlandmarks, hmax, hsize, blind, rlvalue and rldistance. Set multiple heuristics to use multi-gbfs. The default is hadd.
[(--ftp-heuristic | -B) (hadd | hlandmarks | hmax | hsize | blind)]
- ftp deordering – to enable deordering in the ftp planner.
[--ftp-deordering | -d]
- weak equality – enables weak equality in the ftp and ctp planner.
[--weak-equality | -k]
- simultaneity – prunes simultaneous-requiring paths in the ftp and ctp planner.
[--simultaneity | -s]
- ftp-no-backtracking – disables backtracking in ftp planner search.
[--ftp-no-backtracking]
- rlftp model – to indicate the file name of the neural-network to use as heuristic.
[(--rlftp-model | -m) <model-file-name>]
- rlftp gamma – to set the gamma value used in the reinforcement-learning phase. The default is 0.99.
[--rlftp-gamma <gamma-value>]
- rlftp max plan size – to set the max plan size. The default is 1200.
[--rlftp-max-plan-size <max-plan-size-value>]
- rlftp disable use fluents as input – to disable the use of fluents as input in the state vector.
[--rlftp-disable-use-fluents-as-input]
- rlftp disable use actions as input – to disable the use of actions as input in the state vector.
[--rlftp-disable-use-actions-as-input]
- rlftp-disable-use-is-applicable-as-input – to disable the use of ‘is applicable actions’ as input in the state vector.
[--rlftp-disable-use-is-applicable-as-input]
- rlftp-disable-use-was-applied-as-input – to disable the use of ‘was applied actions’ as input in the state vector.
[--rlftp-disable-use-was-applied-as-input]
- rlftp-disable-use-will-be-applicable-as-input – to disable the use of ‘will be applicable actions’ as input in the state vector.
[--rlftp-disable-use-will-be-applicable-as-input]
- rlftp disable use constants as input – to disable the use of constants as input in the state vector.
[--rlftp-disable-use-constants-as-input]
- rlftp disable use goals as input – to disable the use of goals as input in the state vector.
[--rlftp-disable-use-goals-as-input]
- rlftp disable use tn as input – to disable the use of tn as input in the state vector.
[--rlftp-disable-use-tn-as-input]
- rlftp use heuristic as input – to enable the use of the heuristic value as input in the state vector.
[--rlftp-use-heuristic-as-input]
- rlftp use makespan as input – to enable the use of the makespan value as input in the state vector.
[--rlftp-use-makespan-as-input]
transform options
- output – to indicate the output file name where to save the transformed problem.
[--output <output-file>]
validate options
- plan – to indicate the plan file name to validate.
[--plan <plan-file>]
trace-to-plan options
- input-format – to indicate the format of the input trace. The options are smv and uppaal.
[--input-format (smv | uppaal)]
- trace – to indicate the trace input file name.
[--trace <trace-file>]
- output – to indicate the output file name of the generated plan.
[--output <output-file>]
repl options
- disable shell – to disable the interactive shell, allowing only the execution of the input files.
[--disable-shell | -x]
Transformers
- flattener – to flatten the input problem, by expanding all the objects.
- action-grounder – to perform the actions grounding (as much as possible) on the input problem.
- full-grounder – to perform the actions, fluents and constants grounding (as much as possible) on the input problem.
- temporal-uncertainty-compiler – to compile away temporal uncertainty.
- intermediate-effects-compiler – to compile away intermediate effects.
- set-compiler – to compile away sets.
- contains-compiler – to compile away contains expressions.
- partial-order-compiler – to compile away actions with a partial order of time points.
- empty-conditions-compiler – to compile away empty conditions.
- usertype-fluents-compiler – to compile away usertype fluents.
- bounder – to transform the input problem adding the same instances of the provided bounding problem.
- constant-promoter – rewrites the input problem changing fluents that are never assigned after the initial state to constants. This improves planning and grounding performances when it applies.
- free-var-params-compiler – to compile away fluents used as actual parameters of other fluents or constants.
- forall-compiler – to compile away forall expressions.
Tamer REPL
The supported commands are the following.
parse_anml(filename)
– parses the input ANML file and returns aProblemInstance
.parse_pddl(domain_filename, problem_filename)
– parses the input PDDL files and returns aProblemInstance
.parse_ttplan(problem, filename)
– parses the input time-triggered plan file and returns aTTPlan
for the inputProblemInstance
.solve_with_ftp(problem)
– solves the inputProblemInstance
using the ftp planning engine and returns aTTPlan
.solve_with_smt(problem, horizon=100)
– solves the inputProblemInstance
using the smt planning engine and returns aTTPlan
.solve_with_tsimple(problem)
– solves the inputProblemInstance
using the tsimple planning engine and returns aPOTPlan
.validate(problem, ttplan)
– validates the inputTTPlan
for the inputProblemInstance
.flatten_problem(problem)
– flattens the inputProblemInstance
.ground_action_problem(problem, prune_actions=True)
– grounds the actions of the inputProblemInstance
.ground_full_problem(problem, prune_actions=True)
– grounds the inputProblemInstance
.bound_problem(problem, bounding_problem)
– transforms the inputProblemInstance
adding the same instances of the provided boundingProblemInstance
.compile_temporal_uncertainty(problem)
– compiles away temporal uncertainty.compile_intermediate_effects(problem)
– compiles away intermediate effects.compile_usertype_fluents(problem)
– compiles away usertype fluents.set_boolean_option(option, value)
– sets the value of the specified option.set_integer_option(option, value)
– sets the value of the specified option.set_float_option(option, value)
– sets the value of the specified option.set_string_option(option, value)
– sets the value of the specified option.ProblemInstance.is_flat()
– returnsTrue
if theProblemInstance
is flat.ProblemInstance.has_temporal_uncertainty()
– returnsTrue
if theProblemInstance
has temporal uncertainty.ProblemInstance.to_pddl()
– returns a PDDL 2.1 encoding of theProblemInstance
.ProblemInstance.to_ltl()
– returns a LTL encoding of theProblemInstance
.
Common Usage Examples
In the following, we describe some common example usages of Tamer.
- Solve a general ANML problem with SMT-based planner.
$> tamer solve -a smt problem.anml
- Solve a general ANML problem with TSimple-based planner. Note also that this method is unable to solve problem that have required concurrency.
$> tamer solve -a tsimple problem.anml
- Solve a general ANML problem with FTP-based planner.
$> tamer solve -a ftp problem.anml
- Solve a general ANML problem with CTP-based planner.
$> tamer solve -a ctp problem.anml
- Reading and solving a PDDL2.1 problem.
$> tamer solve --pddl domain.pddl problem.pddl
- Since Tamer can perform a number of transformations to the planning problem before actually solving it, it is possible to print the problem passed to the planning engine with the
-p
flag.
$> tamer solve -t full-grounder -p problem.anml
- Convert a problem from PDDL2.1 to ANML.
$> tamer transform --pddl --output problem.anml domain.pddl problem.pddl
- Convert a problem from ANML to PDDL2.1.
$> tamer convert-to --target pddl --domain domain.pddl --problem problem.pddl problem.anml
- Convert a problem from ANML to LTL.
$> tamer convert-to --target ltl --output problem.smv problem.anml
- Convert a problem from ANML to SMV.
$> tamer convert-to --target smv --output problem.smv problem.anml
- Convert a problem from ANML to UPPAAL.
$> tamer convert-to --target uppaal --model model.ta --query query.q problem.anml
- Generate the C++ code of a specialized planner and simulator.
$> tamer code-generate --output-planner gplanner.cxx --output-simulator gsimulator.cxx problem.anml
- Generate the C++ code of a problem free specialized planner.
$> tamer code-generate --output-planner gplanner.cxx -f problem.anml
- Validate a plan.
$> tamer validate --plan plan.txt problem.anml
API
TAMER offers a comprehensive API that allows to interact with the system modeling framework and to the algorithms. From the API it is possible to construct, parse and analyze planning problems, to generate plans and to analyze them. The API is reentrant and allows for multi-threading. Please see the doc/API.md file in the tamer release for more details.