AI Planning with STRIPS

Basic AI planning with STRIPS and PDDL. For details, see the introduction or try the online demo.

Install

npm install strips

Usage

You'll first need to have a domain and problem file in PDDL format. You can create these yourself (see examples) or find them online. The features :strips :typing are supported.

Here is an example that loads a domain and problem from the "Blocks World" domain. This problem involves stacking blocks A, B on one table to a stack AB on another table (where A is on top of B).

Example

var strips = require('strips');

// Load the domain and problem.
strips.load('./examples/blocksworld2/domain.txt', './examples/blocksworld2/problem.txt', function(domain, problem) {
    // Run the problem against the domain.
    var solutions = strips.solve(domain, problem);

    // Display each solution.
    for (var i in solutions) {
        var solution = solutions[i];

        console.log('- Solution found in ' + solution.steps + ' steps!');
        for (var i = 0; i < solution.path.length; i++) {
            console.log((i + 1) + '. ' + solution.path[i]);
        }        
    }
});

Output

- Solution found in 3 steps!
1. move a x y
2. move b x y
3. stack a b y

Methods

load(domainPath, problemPath, callback, isCode)

Loads a domain and problem in PDDL and returns a domain and problem object in the callback. By default, domainPath and problemPath are expected to be file paths, containing PDDL code. To load from strings instead (if you've already loaded the contents of a file), set domainPath and problemPath to the raw PDDL strings and isCode = true.

solve(domain, problem, isDepthFirstSearch = true, maxSolutions = 1, cost = null)

Searches for a solution to the given problem by using depth-first-search, breadth-first-search, or A* search.

The default setting uses depth-first-search and returns the first solution found. To return more solutions, set maxSolutions to a higher value.

To use A* search, provide a function for the "cost" parameter, using the format cost(state), to serve as a heuristic for A*. The function should return an integer, representing the cost of the given state. See starcraft.js for an example.

The cost function may also be passed as the 3rd parameter (ie., solve(domain, problem, cost)). Note, you can also write your own solution algorithm by using the methods below.

graph(domain, problem, minLayers = 0, maxLayers = 0, isSkipNegativeLiterals = false)

Returns a planning graph for a domain and problem.

The planning graph is returned as an array of actions in JSON. In each action, 'precondition' represents parent literals. 'effect' represents child literals. Any action not named 'noop' (no-operation) represents an applicable action, based on the preceding literals. 'noop' is simply a placeholder action to carry each literal forward, from one layer to the next.

Each layer consists of 3-tiers: P0 (literals), A1 (actions), P1 (literals). The format is: P0 = precondition, A1 = actions, P1 = effect.

The planning graph continues adding layers until no new literals and no new actions are discovered. The resulting graph can be used with GraphPlan or other search algorithms. For details on using GraphPlan, see here and here.

If isSkipNegativeLiterals = true, negative literals (mutex) created from an action will be ignored. If you want to include complementary actions in the graph (such as 'Action A on A'), set strips.fast = false.

See example.

getChildStates(domain, state)

Returns an array of all valid child states from a given parent state. Each child state is returned in the format { state: state, action: action }. State is the child state. Action is the applicable action and parameter values on the parent that produced the child state.

applicableActions(domain, state)

Returns a list of applicable actions on the current state. This method uses all possible parameter values and runs each valid action that is defined in the domain against the current state. All actions that satisfy the preconditions are included in the resulting list.

applyAction(action, state)

Applies the action on the state and returns the new (child) state. It is assumed that the action's precondition has already been tested.

isGoal(state, goalState)

Returns true if the state contains the goal state conditions.

isEqual(action1, action2)

Returns true if two actions are equal. Two actions are equal if they contain the same name and parameter values.

stateToString(state)

Converts a JSON state object to a string. Since two states may have the same predicates in different orderings, this method sorts the predicates before returning the string object so they'll always look the same.

actionToString(action)

Converts an action operation to a string. For example: move a b

Settings

strips.fast

Defaults to true, which uses permutationCombination to calculate possible parameter values for actions. Set this property to false to use baseN instead. Using baseN is slower, but will allow strips to utilize all possible solutions. This allows rendering of 'complementary actions', such as 'Action A on A', where normally you would want 'Action A on B'. Changing this setting is only necessary if you are unable to find a solution plan using the default setting.

strips.verbose

Set to true to display status information on the console while searching for a solution.

strips.output

Function to allow redirecting verbose output to different stream. By default: strips.output = function(text) { console.log(text); }

strips.grammarDomainPath

Allows changing the default path to the PEG.js domain grammar file. This file is used to enable parsing of the PDDL domain file.

strips.grammarProblemPath

Allows changing the default path to the PEG.js problem grammar file. This file is used to enable parsing of the PDDL problem file.

Finding Solutions

The default method, strips.solve(...), searches for all solutions using depth-first-search. You can use your own search method by calling the strips methods (listed above) to identify child states and actions yourself. In this way, you can implement A* search, breadth-first-search, or any other type of heuristic solution.

For example, to find all applicable actions for the initial problem state:

var strips = require('strips');
var util = require('util');

// Load the domain and problem.
strips.load('./examples/blocksworld2/domain.txt', './examples/blocksworld2/problem.txt', function(domain, problem) {
	// Get all applicable actions for the initial state.
	var actions = strips.applicableActions(domain, problem.states[0]);
	
	// Display the JSON result.
	console.log(util.inspect(actions, true, 100, true));
});

Similarly, to see the result of applying the first applicable action against the initial state, you can do the following:

// Get all applicable actions for the initial state.
var actions = strips.applicableActions(domain, problem.states[0]);

// Apply first action on the initial state.
var childState = strips.applyAction(actions[0], problem.states[0]);

// Display the JSON result.
console.log(util.inspect(childState, true, 100, true));

Here is a similar example that outputs the state and action to the console:

// Get all applicable actions for the initial state.
var actions = strips.applicableActions(domain, problem.states[0]);

// Display the current state and action that we're going to apply.
console.log('Current state');
console.log(strips.stateToString(problem.states[0]));
console.log('Applying action');
console.log(strips.actionToString(actions[0]));

// Apply first action on the initial state.
var childState = strips.applyAction(actions[0], problem.states[0]);

// Display the resulting modified state.
console.log('New child state');
console.log(strips.stateToString(childState));

Of course, the real power is in designing your own search algorithm using the strips methods. See the default search routine for an idea of how to use the methods to search.

A* Search

A* search works by using a heuristic to guide it down the path of possible moves in the domain. In this manner, it is much faster than simple breadth-first or depth-first search. It will also find an optimal solution that contains the least number of steps.

Strips comes with a built-in A* search algorithm that accepts your own cost function to use as a heuristic. See the section "methods" above. You simply write your own cost function that takes "state" as input and returns an integer as the resulting cost. Here is an example cost function for the starcraft domain to train a marine:

function cost(state) {
    // This is our A* heuristic method to calculate the cost of a state.
    // For Starcraft, the heuristic will be how many required buildings have been built. Subtract x from cost for each correct building, with 0 meaning all required buildings have been made and we're done.
    var cost = 10;

    for (var i in state.actions) {
        var action = state.actions[i].action;

        if (action == 'depot') {
            cost -= 5;
        }
        else if (action == 'barracks') {
            cost -= 5;
        }
    }
    
    return cost;
}

You can also implement your own A* search to find a solution. Since the strips library exposes its internal methods, you can implement your own search algorithm. The core idea to a custom search method is to use the strips methods isGoal() and getChildStates() to iterate through all states and actions. Once you have a list of child states, apply your heuristic to calculate a cost for each state. Then sort the states by cost so that A* can choose the next cheapest state to move to. You can see the details in the solve() methods in strips.

Have fun!

License

MIT

Copyright (c) 2015 Kory Becker http://primaryobjects.com/kory-becker