[nit.git] / lib / ai / search.nit

# This file is part of NIT ( http://www.nitlanguage.org ).
#
# This file is free software, which comes along with NIT.  This software is
# distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY;
# without  even  the implied warranty of  MERCHANTABILITY or  FITNESS FOR A
# PARTICULAR PURPOSE.  You can modify it is you want,  provided this header
# is kept unaltered, and a notification of the changes is added.
# You  are  allowed  to  redistribute it and sell it, alone or is a part of
# another product.

# Basic framework for search problems and solver.
#
# The module provides a simple abstract class `SearchProblem[S,A]` to be specialized for a specific problem.
#
# The concrete class `SearchSolver` is used to configure, query, and run a solver for a given problem.
module search

import realtime
import trees

# Abstract search problem over immutable states (`S`) and actions (`A`).
#
# This class serves to model problems of planing and path-finding.
# A state, `S`, is a point in the search problem and fully models a given state of the world.
# An action, `A`, is an available mean of transition between two states.
#
# This abstract class is generic made to work with any kind of states and actions.
# Therefore, specific subclasses must be developed to implement the required services:
#
# * `initial_state`
# * `actions`
# * `apply_action`
# * `is_goal`
#
# Note that the implemented methods should not temper with the parameters since it is expected
# that they remain unmodified.
#
#
# # Basic search
#
# These tree method are enough to trigger a basic search.
#
# The methods `breadth_first` and `depth_first` return pre-configured solvers for, respectively,
# a breadth-first search, a depth-first search.
#
#
# # Advanced search
#
# The `cost` method can be implemented to represent the cost of a single transition.
# By default, the cost is 1.
#
# The `heuristic` method can be implemented to represent a lower-bound estimation of the remaining
# cost to reach a goal state.
# By default, the heuristic is 0.
#
# If one of these (or both) methods are implemented, the `astar` method will return a pre-configured
# solver for a A* search.
#
# More configuration and optimization on the search can to be done in the `SearchSolver` class.
interface SearchProblem[S: Object, A]
        # The starting state for the problem
        fun initial_state: S is abstract

        # The available applicable actions for a given state.
        # While there is a potential large number of distinct states and actions, there should be only
        # a small number of possible action from a specific state (a small, or at least finite, branching factor).
        fun actions(state: S): nullable SequenceRead[A] is abstract

        # The new state when applying a given action
        #
        # The returned state can be:
        # * a completely new state,
        # * an existing state,
        # * a new state but equals to an existing state
        #   in this case, ensure that the `==` and `hash` method
        #   are correctly implemented.
        #
        # Remember, this method should not modify its parameters.
        #
        # REQUIRE: `actions(state).has(action)`
        fun apply_action(state:S, action:A): S is abstract

        # Is the state a goal state?
        # Once a goal state is found, the solver is automatically stopped.
        # A problem can have 0, one or more goals if it makes sense
        # but solver must be used accordingly.
        # Default: no goals
        fun is_goal(state:S): Bool do return false

        # The cost for `action` from `old_state` to `new_state`
        # REQUIRE: `apply_action(old_state, action) == new_state`
        # Default: `1`.
        # Note that having a 0 or negative value can make some search
        # algorithms fail, or not terminate.
        fun cost(state:S, action:A, state2: S): Float do return 1.0

        # An heuristic of the estimated `cost` going from `state` to a goal state.
        #
        # Is is expected that the heuristic is *admissible*, it means its is an
        # optimistic estimation that never an over-estimate, thus is cannot be#
        # higher than the lowest possible remaining cost.
        # See `SearchSolver::do_revisit` for details.
        #
        # Default: `0`.
        fun heuristic(state: S): Float do return 0.0

        # return a new breadth-first solver
        fun breadth_first: SearchSolver[S, A]
        do
                var todo = (new Array[SearchNode[S, A]]).as_fifo
                var sol = new SearchSolver[S, A](self, todo)
                return sol
        end

        # return a new depth-first solver
        fun depth_first: SearchSolver[S, A]
        do
                var todo = (new List[SearchNode[S, A]]).as_lifo
                var sol = new SearchSolver[S, A](self, todo)
                return sol
        end

        # return a new best-first solver
        #
        # notes:
        # * if `heuristic` is not defined, this is basically a Dijkstra search.
        # * if `cost` in not defined either, this is basically a breadth-first search.
        fun astar: SearchSolver[S, A]
        do
                var cpt = new NodeComparator[S, A]
                var todo = new MinHeap[SearchNode[S, A]](cpt)
                var sol = new SearchSolver[S, A](self, todo)
                return sol
        end

        # Create the initial node in the search-tree.
        # Used internally by the solvers but made public for those who want to replay a plan.
        fun initial_node: SearchNode[S, A]
        do
                var res = new SearchNode[S,A](self, initial_state, null, null, 0.0, 0)
                res.compute_heuristic
                return res
        end

        # Negligible quantity for float comparisons
        # Because of float imprecision, two really near float values should be considered equals.
        # However, the specific epsilon value could be specific to the problem.
        #
        # The epsilon value is used for cost comparisons.
        #
        # Default: 1E-9
        fun epsilon: Float do return 0.000000001

        # Run and evaluate solvers with multiple configuration.
        # This method can be used to evaluate which configuration to choose by default for a given problem.
        #
        # `steps` is the maximum number of steps a giver configuration can run.
        fun run_configs(steps: Int)
        do
                var s

                var c = 0
                loop
                        if astar.run_config(steps, c, "A*") then break
                        c += 1
                end
        end

        # Various Map implementations of memory.
        # In order to try more configurations with `run_config`, this method
        # is called to provide alternative implementation.
        #
        # For instance, a subclass can redefine this method and extends the result with an additional `RBTreeMap`.
        # Note: because the true nature of the sates `S` is left to the user, some
        # specific Map implementation could be more efficient than a HashMop.
        #
        # Default: A `HashMap`
        fun make_memory: Array[Map[S, SearchNode[S, A]]]
        do
                var res = new Array[Map[S, SearchNode[S, A]]]
                res.add new HashMap[S, SearchNode[S, A]]
                return res
        end
end

# A running solver for a given problem, to configure and control.
#
# For a given problem, a lot of variation of search algorithms can be made.
# Thus this class permit the user to control the parameters of the search algorithm.
#
# Note that this class is not meant to be specialized, and usually not instantiated directly.
#
#
# # Basic run and result.
#
# 1. Instantiate it with the method `breadth_first`, `depth_first`, or `astar` from `SearchProblem`.
# 2. Apply the method `run`, that will search and return a solution.
# 3. Retrieve information from the solution.
#
# ~~~~
# var p: SearchProblem = new MyProblem
# var res = p.astar.run
# if res != null then print "Found plan with {res.depth} actions, that cost {res.cost}: {res.plan.join(", ")}"
# ~~~~
#
#
# # Step-by-step runs and multiple runs
#
# The `run_steps` method (see also `steps`, and `steps_limit`) can be used to run only a maximum number of steps.
# This method can be used as a *co-routine* and run them periodically for a small amount of time.
#
# `run` and `run_steps` return the next solution.
# A subsequent call to `run` returns the following solution and so on.
#
# When there is no more solutions available, `is_running` become false.
#
#
# # Search-trees
#
# Internally, solvers use a search-tree where nodes are labeled with states, and edges are labeled with actions.
# See `SearchNode` for details.
#
# The `run` method return a `SearchNode` that can be used to retrieve a lot of useful information,
# like the full `path` or the `plan`.
#
#
# # Configuration
#
# The solver provide some *knobs* to control how the search-tree is visited.
#
# * `memorize` (see also `memorize_late`)
# * `do_revisit` (see also `revisits`)
# * `depth_limit` (see also `iterative_deepening` and `depth_limit_reached`)
class SearchSolver[S: Object, A]
        # The problem currently solved
        var problem: SearchProblem[S, A]

        # The currently open nodes to process.
        # They are the open nodes.
        #
        # It is the nature of the queue that control how the solver works.
        # However, the methods `SearchProblem::breadth_first`, `SearchProblem::depth_first`,
        # and `SearchProblem::astar` takes care of its correct initialization.
        private var todo: Queue[SearchNode[S, A]]

        # Is the solver still running?
        # A running solver has not yet exhausted all the possible solutions.
        var is_running: Bool = true

        # Does the solver need to memorize visited states?
        # When activated, there is an increased memory consumption since
        # all visited states must be kept in memory,
        # However, there is real a gain, since visited nodes are not
        # revisited (unless needed by `do_revisit`)
        #
        # Default: `true`
        #
        # Note: If the graph of states has circuits, then a memory-less search may not terminate.
        var memorize: Bool = true is writable

        # Use memory only on visited (closed) state.
        # Less memory operations, but two big drawbacks:
        # * duplicated nodes can fill the `todo` queue (and the memory)
        # * duplicated nodes require more invocation of `SearchProblem::heuristic`
        #
        # Note: if `memorize` is false, then this has no effect.
        #
        # Default: `false`
        var memorize_late: Bool = false is writable

        # Storage of nodes when `memorize` is activated
        # Each state is associated with a node.
        # This permit:
        #   * to avoid to revisit visited nodes (see `do_revisit`)
        #   * to avoid to reinvoke `heuristic` on known states (unless `memorize_late` is set)
        private var memory: Map[S, SearchNode[S, A]] = new HashMap[S, SearchNode[S, A]]

        # Total number of time an already memorized node is seen again.
        # If `memorize_late` is set, then only visited nodes are counted.
        # Otherwise, nodes in the todo-list are also considered.
        var memorized = 0

        # Revisit states when a better path to them is found.
        # Such revisits generally triggers more revisits because they yield
        # better path to their neighbors.
        #
        # If `false`, visited states are never revisited.
        #
        # With astar and an admissible heuristic, no visited node should be revisited.
        # If the heuristic is not admissible, one may consider set this to `true`.
        #
        # Obviously, if `memorize` is false, then the value has no specific effect
        # since all states are considered unvisited.
        #
        # Default: `false`.
        #
        # See also `revisits` and `SearchNode::revisits`.
        var do_revisit: Bool = false is writable

        # Total number of states (potentially) revisited.
        #
        # It is the number of time that a better path to a visited state is found.
        # With astar and a really admissible heuristic, this number should stay 0.
        # So check this value if you are not sure of the heuristic.
        #
        # Note that states are effectively revisited if `do_revisit` is activated.
        var revisits = 0

        # The solution found by the last `run`.
        #
        # ensure `solution != null implies problem.is_goal(solution.state)`
        var solution: nullable SearchNode[S,A] = null

        # Nearest solution found (up to date).
        # The nearest solution is the one with the smallest heuristic value.
        # The cost is not considered.
        var nearest_solution: nullable SearchNode[S,A] = null

        # Limit in the depth search.
        #
        # States found above this limit are not considered.
        #
        # Use 0 for no limit.
        # Default: 0
        # See also: `iterative_deepening`
        var depth_limit: Int = 0 is writable

        # How much time a `depth_limit` was reached?
        #
        # This can be used to query if some solutions may have been
        # ignored because of a `depth_limit`.
        #
        # This is also used automatically if `iterative_deepening` is activated.
        var depth_limit_reached: Int = 0

        # Increase amount for an iterative deepening search.
        # It =0, then the iterative deepening search is disabled.
        # If >0, then `depth_limit` is automatically increased when the todo
        # queue is empty but the `depth_limit` was reached in the previous iteration.
        # Default: 0
        var iterative_deepening: Int = 0

        # The total steps executed since the beginning
        # A step is the visit of a node in the `todo`-list
        var steps: Int = 0

        # The total number of nodes created
        var nodes = 0

        # Limit in the number of steps for a `run`.
        #
        # One can modify this value then `run` or just call `run_steps`.
        #
        # Use 0 for no limit.
        # Default: 0
        var steps_limit: Int = 0 is writable

        # Total number of neighbors considered.
        var neighbors = 0

        # The average number of neighbors by nodes.
        fun branching_factor: Float do return neighbors.to_f / steps.to_f

        # Update `steps_limit` then just run some additional steps
        # Return the best solution so far (if any)
        fun run_steps(steps: Int): nullable SearchNode[S,A]
        do
                assert steps > 0
                self.steps_limit = self.steps + steps
                return run
        end

        # Reset the search from the initial state.
        # Is used at the beginning and with `iterative_deepening`.
        private fun start
        do
                assert todo.is_empty
                depth_limit_reached = 0
                var initial_node = problem.initial_node
                if memorize and not memorize_late then memory[initial_node.state] = initial_node
                initial_node.id = nodes
                nodes += 1
                todo.add initial_node
        end

        # Run the solver and return the next solution (if any)
        # Return null is one of these is true:
        # * `steps_limit` is reached
        # * the `todo` queue is empty (eg. no reachable solution)
        fun run: nullable SearchNode[S,A]
        do
                if steps == 0 then start

                var nearest = nearest_solution
                loop
                        # Enough work
                        if steps_limit > 0 and steps >= steps_limit then break

                        #print "todo={todo.length}"
                        #print "  {todo.join("\n  ")}"

                        # Next node, please
                        if todo.is_empty then
                                # iterative depth search?
                                if depth_limit <= 0 or iterative_deepening <= 0 or depth_limit_reached == 0 then
                                        is_running = false
                                        break
                                end

                                depth_limit += iterative_deepening
                                start
                        end
                        var node = todo.take

                        # Skip `old` stuff
                        # Because `Queue` has no remove :(
                        if node.drop then continue

                        var state = node.state

                        if memorize and memorize_late then
                                # Is the state already visited?
                                var old = memory.get_or_null(state)
                                if old != null then
                                        memorized += 1
                                        if old.cost - node.cost < problem.epsilon then continue
                                        revisits += 1
                                        if not do_revisit then
                                                old.revisits += 1
                                                continue
                                        end
                                        node.revisits = old.revisits + 1
                                end
                                memory[state] = node
                        end

                        steps += 1
                        assert node.steps == 0
                        node.steps = steps
                        self.node = node

                        # Keep trace to the nearest
                        if nearest == null or node.heuristic < nearest.heuristic then
                                nearest = node
                                nearest_solution = node
                        end

                        #print "try {node}"
                        #print "try {node}; todo={todo.length}"

                        # Won?
                        if problem.is_goal(state) then
                                solution = node
                                return node
                        end

                        # Ignore successors states if the depth limit is reached
                        if depth_limit > 0 and node.depth >= depth_limit then
                                depth_limit_reached += 1
                                continue
                        end

                        # Now, expand!
                        var actions = problem.actions(state)
                        if actions == null then continue
                        for action in actions do
                                neighbors += 1

                                # Fast track if no memory or late memory
                                if not memorize or memorize_late then
                                        var new_node = node.apply_action(action)
                                        new_node.id = nodes
                                        nodes += 1
                                        todo.add(new_node)
                                        continue
                                end

                                # Get the state and the cost. Do not create the node yet.
                                var new_state = problem.apply_action(state, action)
                                var new_cost = node.cost + problem.cost(state, action, new_state)

                                # So check if the state was already seen
                                var old = memory.get_or_null(new_state)
                                if old != null then
                                        memorized += 1
                                        # If not better, then skip
                                        if old.cost - new_cost < problem.epsilon then continue
                                        # If visited and do not revisit, then skip
                                        if old.steps > 0 and not do_revisit then
                                                old.revisits += 1
                                                revisits += 1
                                                continue
                                        end
                                        # Even if `==`, reuse the same state object so
                                        # * it may helps the GC
                                        # * user-cached things in the previous state can be reused
                                        new_state = old.state
                                end

                                # Finally, create the node
                                var new_node = new SearchNode[S, A](problem, new_state, node, action, new_cost, node.depth+1)
                                new_node.id = nodes
                                nodes += 1

                                if old == null then
                                        # Compute heuristic and cost
                                        new_node.compute_heuristic
                                else
                                        # Reuse heuristic and update the cost
                                        var h = old.heuristic
                                        new_node.heuristic = h
                                        new_node.score = new_cost + h

                                        # Is `old` a visited node?
                                        if old.steps == 0 then
                                                # Old is still in the todo list, so drop it
                                                old.drop = true
                                        else
                                                # Old was visited, so revisit it
                                                new_node.revisits = old.revisits + 1
                                                revisits += 1
                                                #print "found {old.cost}>{new_cost}:{old.cost>new_cost} d={old.cost-new_cost}\n\t{old}\nthat is worse than\n\t{new_node}"
                                        end
                                end
                                memory[new_state] = new_node

                                todo.add(new_node)
                        end
                end
                return null
        end

        # The last visited node.
        # Unless when debugging, the last visited node is not really meaningful.
        var node: nullable SearchNode[S, A] = null

        redef fun to_s
        do
                var res ="steps={steps} nodes={nodes} todo={todo.length}"
                if neighbors > 0 then res += " n={neighbors} (bf={branching_factor})"
                if revisits > 0 then res += " re={revisits}"
                if memorized > 0 then res += " mem={memorized}"
                var n = solution
                if n != null then
                        res += " last={n}"
                else
                        n = nearest_solution
                        if n != null then res += " near={n}"
                end
                return res
        end

        # Run the configuration number `i`, for `steps` number of steps.
        # The message `msg` suffixed by the name of the configuration is printed followed by the result
        #
        # This method is used by `SearchProblem::run_configs`
        fun run_config(steps: Int, i: Int, msg: String): Bool
        do
                do
                        if i == 0 then
                                msg += " -mem"
                                memorize = false
                                break
                        end
                        i -= 1

                        var mems = problem.make_memory
                        memory = mems[i % mems.length]
                        msg += " {memory.class_name}"
                        i = i / mems.length

                        if i % 2 == 0 then
                                msg += " +mem"
                                memorize = true
                                memorize_late = false
                        else
                                msg += " +mem_late"
                                memorize = true
                                memorize_late = true
                        end
                        i = i / 2

                        if i % 2 == 0 then
                                msg += " +revisit"
                                do_revisit = true
                        else
                                msg += " -revisit"
                                do_revisit = false
                        end
                        i = i / 2

                        if i >= 1 then return true

                end
                print msg

                var t = new Clock
                var res = run_steps(steps)
                print "\t{self}"
                var l = t.lapse
                print "\ttime={l}"
                return false
        end
end

# Used to compare nodes with their score.
# Smaller is score, smaller is the node.
private class NodeComparator[S: Object, A]
        super Comparator[SearchNode[S, A]]
        redef fun compare(a,b) do return a.score <=> b.score
end

# A node in the search-tree visited by a `SearchSolver`.
# In search-trees, nodes are labeled with states (`S`), and edges by actions (`A`).
#
# The root node is labeled by the initial state of the problem.
#
# This class is exposed to allow queries on the solution provided by the solver.
class SearchNode[S: Object, A]
        # A flag that indicate that `self` is virtually removed from the todo-list.
        # `self` was added to the todo-list but that a better node for the
        # same state was found latter.
        private var drop = false

        # The associated problem
        var problem: SearchProblem[S, A]

        # The state associated to `self`.
        # The state labels the node `self`.
        var state: S

        # Is `self` a root node of the search-tree?
        # ensure: `result` == `parent == null` and `result`== `action == null`
        fun is_root: Bool do return parent == null

        # The previous node in the search-tree (if not root).
        var parent: nullable SearchNode[S, A]

        # The action used to go from `parent` to `self` (if not root)
        # The action labels the edge from `parent` to `self`.
        var action: nullable A

        # The past cost (g) from the root node to `self`.
        var cost: Float

        # The heuristic from self to the goal (according to `problem.heuristic(state)`
        # It is the future cost (h)
        var heuristic: Float is noinit

        # The sum of `cost` and `heuristic`
        # It is the f function.
        var score: Float is noinit

        # Update `heuristic` and `score` according to `problem`.
        private fun compute_heuristic
        do
                var h = problem.heuristic(state)
                heuristic = h
                score = cost + h
        end

        # The depth of `self` in the search tree
        # It is the number of parents to the root node.
        var depth: Int

        # The number of steps needed by the solver to process `self`
        # It is just a useless generation number, but could be used to evaluate
        # the behavior of search algorithms.
        var steps: Int = 0

        # The rank of creation of nodes by the solver.
        # It is just a useless generation number, but could be used to evaluate
        # the behavior of search algorithms.
        var id: Int = 0

        # The number of (potential) revisits of `node`.
        # This information can be used to debug search algorithms.
        # And to detect when heuristics are not admissible.
        #
        # See `SearchSolver::revisits` and `SearchSolver::do_revisit`
        # for details.
        var revisits: Int = 0

        # Create a new child node for the next state, according to `problem`.
        # Used internally by the solvers but made public for those who want to replay a plan.
        #
        # ensure `result.parent == self`
        # ensure `result.action == action`
        fun apply_action(action: A): SearchNode[S, A]
        do
                var new_state = problem.apply_action(state, action)
                var new_cost = problem.cost(state, action, new_state)
                var new_node = new SearchNode[S, A](problem, new_state, self, action, cost + new_cost, depth+1)
                new_node.compute_heuristic
                return new_node
        end

        # Build the sequence of nodes from the initial node to `self`
        #
        # ensure `result.first.is_root and result.last == self`
        fun path: Sequence[SearchNode[S, A]]
        do
                var res = new List[SearchNode[S, A]]
                res.add(self)
                var node = parent
                while node != null do
                        res.unshift(node)
                        node = node.parent
                end
                return res
        end

        # Build a sequence of actions from the initial state to `self`
        # See `path` for a more detailed plan.
        fun plan: Sequence[A]
        do
                var res = new List[A]
                var node: nullable SearchNode[S, A] = self
                while node != null do
                        var a = node.action
                        if a != null then res.unshift(a)
                        node = node.parent
                end
                return res
        end

        # Just print a detailed path on the screen
        fun dump
        do
                print "result:{state}"
                for n in path do
                        var a = n.action
                        if a != null then print "    + {a or else ""}"
                        print "  {n.steps}: {n.state} ({n.cost}$)"
                end
        end

        redef fun to_s do return "#{steps}/{id} d={depth} f={cost+heuristic} g={cost} h={heuristic}: {state}"
        #redef fun to_s do return "#{steps} f={(cost+heuristic).to_i} g={cost.to_i} h={heuristic.to_i}"
end