Basic implementation of A* in Erlang
3 min read

Basic implementation of A* in Erlang

Recently I had to write some path finding algorithms in Erlang. The first version I chose was A*. But, there is no easy way to implent A* in a distributed way. So, this is the simplest implementation possible. I may rewrite it later if I find a better way.

This code is mostly a modified version of this one.

The code hosted on gist follows below, followed by some notes.

-module(astar).

-type cnode() :: {integer(), integer()}.

-define(MINX, 0).
-define(MINY, 0).
-define(MAXX, 10).
-define(MAXY, 10).

-export([
         astar/2,
         neighbour_nodes/2
        ]).

%% @doc Performs A* for finding a path from `Start' node to `Goal' node
-spec astar(cnode(), cnode()) -> list(cnode()) | failure.
astar(Start, Goal) ->
    ClosedSet = sets:new(),
    OpenSet   = sets:add_element(Start, sets:new()),

    Fscore    = dict:store(Start, h_score(Start, Goal), dict:new()),
    Gscore    = dict:store(Start, 0, dict:new()),

    CameFrom  = dict:store(Start, none, dict:new()),

    astar_step(Goal, ClosedSet, OpenSet, Fscore, Gscore, CameFrom).

%% @doc Performs a step of A*.
%% Takes the best element from `OpenSet', evaluates neighbours, updates scores, etc..
-spec astar_step(cnode(), set(), set(), dict(), dict(), dict()) -> list(cnode()) | failure.
astar_step(Goal, ClosedSet, OpenSet, Fscore, Gscore, CameFrom) ->
    case sets:size(OpenSet) of
        0 ->
            failure;
        _ ->
            BestStep = best_step(sets:to_list(OpenSet), Fscore, none, infinity),
            if
                Goal == BestStep ->
                    lists:reverse(reconstruct_path(CameFrom, BestStep));
                true ->
                    Parent     = dict:fetch(BestStep, CameFrom),
                    NextOpen   = sets:del_element(BestStep, OpenSet),
                    NextClosed = sets:add_element(BestStep, ClosedSet),
                    Neighbours = neighbour_nodes(BestStep, Parent),

                    {NewOpen, NewF, NewG, NewFrom} = scan(Goal, BestStep, Neighbours, NextOpen, NextClosed, Fscore, Gscore, CameFrom),
                    astar_step(Goal, NextClosed, NewOpen, NewF, NewG, NewFrom)
            end
    end.

%% @doc Returns the heuristic score from `Current' node to `Goal' node
-spec h_score(Current :: cnode(), Goal :: cnode()) -> Hscore :: number().
h_score(Current, Goal) ->
    dist_between(Current, Goal).

%% @doc Returns the distance from `Current' node to `Goal' node
-spec dist_between(cnode(), cnode()) -> Distance :: number().
dist_between(Current, Goal) ->
    {X1, Y1} = Current,
    {X2, Y2} = Goal,
    abs((X2-X1)) + abs((Y2-Y1)).

%% @doc Returns the best next step from `OpenSetAsList'
%% TODO: May be optimized by making OpenSet an ordered set.
-spec best_step(OpenSetAsList :: list(cnode()), Fscore :: dict(), BestNodeTillNow :: cnode() | none, BestCostTillNow :: number() | infinity) -> cnode().
best_step([H|Open], Score, none, infinity) ->
    V = dict:fetch(H, Score),
    best_step(Open, Score, H, V);

best_step([], _Score, Best, _BestValue) ->
    Best;

best_step([H|Open], Score, Best, BestValue) ->
    Value = dict:fetch(H, Score),
    case Value < BestValue of
        true ->
            best_step(Open, Score, H, Value);
        false ->
            best_step(Open, Score, Best, BestValue)
    end.

%% @doc Returns the neighbour nodes of `Node', and excluding its `Parent'.
-spec neighbour_nodes(cnode(), cnode() | none) -> list(cnode()).
neighbour_nodes(Node, Parent) ->
    {X, Y} = Node,
    [
     {XX, YY} ||
     {XX, YY} <- [{X-1, Y}, {X, Y-1}, {X+1, Y}, {X, Y+1}],
     {XX, YY} =/= Parent,
     XX >= ?MINX,
     YY >= ?MINY,
     XX =< ?MAXX,
     YY =< ?MAXY
    ].

%% @doc Scans the `Neighbours' of `BestStep', and adds/updates the Scores and CameFrom dicts accordingly.
-spec scan(
        Goal :: cnode(),
        BestStep :: cnode(),
        Neighbours :: list(cnode()),
        NextOpen :: set(),
        NextClosed :: set(),
        Fscore :: dict(),
        Gscore :: dict(),
        CameFrom :: dict()
       ) ->
    {NewOpen :: set(), NewF :: dict(), NewG :: dict(), NewFrom :: dict()}.
scan(_Goal, _X, [], Open, _Closed, F, G, From) ->
    {Open, F, G, From};
scan(Goal, X, [Y|N], Open, Closed, F, G, From) ->
    case sets:is_element(Y, Closed) of
        true ->
            scan(Goal, X, N, Open, Closed, F, G, From);
        false ->
            G0 = dict:fetch(X, G),
            TrialG = G0 + dist_between(X, Y),
            case sets:is_element(Y, Open) of
                true ->
                    OldG = dict:fetch(Y, G),
                    case TrialG < OldG of
                        true ->
                            NewFrom = dict:store(Y, X, From),
                            NewG    = dict:store(Y, TrialG, G),
                            NewF    = dict:store(Y, TrialG + h_score(Y, Goal), F), % Estimated total distance from start to goal through y.
                            scan(Goal, X, N, Open, Closed, NewF, NewG, NewFrom);
                        false ->
                            scan(Goal, X, N, Open, Closed, F, G, From)
                    end;
                false ->
                    NewOpen = sets:add_element(Y, Open),
                    NewFrom = dict:store(Y, X, From),
                    NewG    = dict:store(Y, TrialG, G),
                    NewF    = dict:store(Y, TrialG + h_score(Y, Goal), F), % Estimated total distance from start to goal through y.
                    scan(Goal, X, N, NewOpen, Closed, NewF, NewG, NewFrom)
            end
    end.

%% @doc Reconstructs the calculated path using the `CameFrom' dict
-spec reconstruct_path(dict(), cnode()) -> list(cnode()).
reconstruct_path(CameFrom, Node) ->
    case dict:fetch(Node, CameFrom) of
        none ->
            [Node];
        Value ->
            [Node | reconstruct_path(CameFrom, Value)]
    end.

Notes

  • Variables MINX, MINY, MAXX and MAXY can be modified to increase the size of the map. The function neighbour_nodes/2 can be modified to add obstacles.

  • To test, enter in erlang shell:

c(astar).
astar:astar({1, 1}, {10, 10}).
  • The cnode() structure represents some sort of coordinate. To use some other structure, the functions neighbour_nodes/2, h_score/2, and distance_between/2 have to be modified for the new structure.

  • The current heuristic does not penalize for turns, so the resultant path tends to follow a diagonal looking shape. For correcting this, either diagonal movements can be allowed (by modifying the neighbours function), or turning could be penalized in the heuristic function (current direction would have to be tracked).

Follow me on Twitter