Conference PaperPDF Available

Rolling Horizon Evolution versus Tree Search for Navigation in Single-Player Real-Time Games


Abstract and Figures

In real-time games, agents have limited time to respond to environmental cues. This requires either a policy defined up-front or, if one has access to a generative model, a very efficient rolling horizon search. In this paper, different search techniques are compared in a simple, yet interesting, real- time game known as the Physical Travelling Salesman Problem (PTSP). We introduce a rolling horizon version of a simple evolutionary algorithm that handles macro-actions and compare it against Monte Carlo Tree Search (MCTS), an approach known to perform well in practice, as well as random search. The experimental setup employs a variety of settings for both the action space of the agent as well as the algorithms used. We show that MCTS is able to handle very fine-grained searches whereas evolution performs better as we move to coarser-grained actions; the choice of algorithm becomes irrelevant if the actions are even more coarse-grained. We conclude that evolutionary algorithms can be a viable and competitive alternative to MCTS.
Content may be subject to copyright.
Rolling Horizon Evolution versus Tree Search for
Navigation in Single-Player Real-Time Games
Diego Perez
University of Essex
Colchester CO4 3SQ
United Kingdom
Spyridon Samothrakis
University of Essex
Colchester CO4 3SQ
United Kingdom
Simon M. Lucas
University of Essex
Colchester CO4 3SQ
United Kingdom
Philipp Rohlfshagen
University of Essex
Colchester CO4 3SQ
United Kingdom
In real-time games, agents have limited time to respond to
environmental cues. This requires either a policy defined
up-front or, if one has access to a generative model, a very
efficient rolling horizon search. In this paper, different search
techniques are compared in a simple, yet interesting, real-
time game known as the Physical Travelling Salesman Prob-
lem (PTSP). We introduce a rolling horizon version of a sim-
ple evolutionary algorithm that handles macro-actions and
compare it against Monte Carlo Tree Search (MCTS), an
approach known to perform well in practice, as well as ran-
dom search. The experimental setup employs a variety of
settings for both the action space of the agent as well as
the algorithms used. We show that MCTS is able to han-
dle very fine-grained searches whereas evolution performs
better as we move to coarser-grained actions; the choice of
algorithm becomes irrelevant if the actions are even more
coarse-grained. We conclude that evolutionary algorithms
can be a viable and competitive alternative to MCTS.
Categories and Subject Descriptors
I.2 [Computing Methodologies]: Artificial Intelligence—
ai, games, real time, evolutionary algorithms, mcts
Traditionally, research in Game AI has focused on 2-player
perfect information games where the time allowed to make
a move is measured in seconds or minutes. Lately, research
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. Tocopy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
GECCO’13, July 6–10, 2013, Amsterdam, The Netherlands.
Copyright 2013 ACM 978-1-4503-1963-8/13/07 ...$15.00.
has started to focus on more complex domains such as those
where the whole state of the game is unobserved by the play-
ers (as in many card games) or is affected by random events
(e.g. in stochastic games like Backgammon). Likewise, real-
time constraints have attracted increasing attention, with
particular emphasis on video games, where moves need to
be made every number of milliseconds. In this paper we will
concentrate on the real-time aspect of a game, which requires
a search algorithm to converge very fast without much over-
head. It is therefore worthwhile to investigate how search
methods are able to explore the search space of those games
characterised by limited decision time.
The benchmark (and the experimental testbed) used in
this paper is the Physical Travelling Salesman Problem (or
PTSP), a deterministic single player game that consists of a
collection of several waypoints scattered around a map full
of obstacles. The challenges here include finding the best
sequence of waypoints, as well as the challenge of navigat-
ing the player to visit each waypoint in sequence in as little
time as possible. This paper focuses in the navigation as-
pect of the problem. As a real-time game, PTSP provides
a small time budget to decide the next move of the player.
Additionally, the game allows the player to run simulations
in order to foresee the outcome of applying any of the avail-
able actions. However, the final state of the game is usually
never reached, as the number of moves required to do so by
far exceeds the time available to execute them.
As a single-player game, the PTSP allows players that
have a model of the game to predict the rewards of the ac-
tions taken without any ambiguity. This feature allows the
controllers to create precise macro-actions (or repetitions of
the same single actions) to reach further states when evalu-
ating the next move to make.
The PTSP has been selected because it holds certain simi-
larities with some modern video-games, at a simplified level.
For instance, real-time games need an action to be supplied
within a few milliseconds, providing a very limited budget to
plan the next move. Furthermore, as in the PTSP, the du-
ration of the game is long enough so the agent is not able to
foresee all the outcomes from the moves made. Finally, the
fact that the PTSP has featured in international conferences
as a competition (see Section 3) allows for a direct compar-
ison between the approaches shown here and the ones from
the contest.
Four different algorithms are presented in this paper: two
versions of a Rolling Horizon Genetic Algorithm (GA and
GAC), Rolling Horizon Random Search (RS) and Monte
Carlo Tree Search (MCTS). Each one of these algorithms
follows a similar scheme in order to solve the problem: first,
the order of waypoints to visit is obtained by a common TSP
solver. Then, a macro-action handler performs the desired
moves while it requests future actions from the appropriate
algorithm. Each algorithm performs the search differently,
but uses the same value function that evaluates a given state
of the game. Therefore, the focus of this paper is on the
ability of the algorithm to explore the search space, using
different macro-action lengths.
This paper is organised as follows. Section 2 discusses
tree search and evolutionary algorithms in real-time games.
Section 3 describes the PTSP and Section 4 introduces the
algorithms and controllers used to play the game. Section 5
details the experimental setup and discusses the results ob-
tained. The paper is concluded in Section 6, where some
possible extensions of this work are discussed.
2.1 Monte Carlo Tree Search
MCTS is a tree search technique that has been widely
used in board games, particularly in the game of Go. Go
is a territorial board game played in a 19 ×19 square grid,
in which the two players aim to surround their opponent’s
stones. The game of Go is considered the drosophila of Game
AI, and MCTS players have achieved professional level play
in the reduced version of the game (9 ×9 board size) [8].
MCTS is considered to be anytime: the algorithm is able
to provide a valid solution (next move in the game) at any
moment in time, independently from the number of itera-
tions that have been performed (although more iterations
generally produce better results). In contrast, other algo-
rithms (such A*) usually require to have finished to provide
an initial action to execute. For this reason, MCTS is a
suitable method for real-time domains, and it has been em-
ployed extensively in the literature.
A popular real-time game that has been employed as a
benchmark for using MCTS is Ms. Pac Man. In this game,
the user controls the Pac Man agent with the objective of
clearing the maze and eating all the pills without being cap-
tured by the ghosts. This game, similar to the PTSP, is
open ended (i.e. the end game state is usually not reached
within the simulations). Robles et al [16] include hand-coded
heuristics to guide the simulations towards more promising
portions of the search space. Other authors also include do-
main knowledge from the game to bias the search in MCTS,
such as [18, 7].
Regarding single player games, MCTS has been used in
several puzzles. Examples are SameGame [9], where the
player aims to destroy contiguous items of the same colour
distributed in a rectangular grid, or Morpion Solitaire [4], a
connection puzzle where the objective is to join vertices of
a graph with lines that contain at least five nodes. PTSP
itself has been addressed with MCTS techniques, as Perez
et al. [11] did to show the importance of adding heuristics to
the Monte Carlo (MC) simulations in real-time domains, or
Powley et al. [14], who describe the MCTS algorithm that
won them the first edition of the PTSP competition.
An extensive survey of Monte Carlo Tree Search methods
has been written by Browne et al. [2], where the algorithm,
variations and applications are described. It is worth noting
that Monte Carlo Tree Search, when combined with UCB1
(see later in this paper) reaches asymptotically logarithmic
regret [3]. One can intuitively think of regret as“opportunity
2.2 Rolling Horizon Evolutionary Algorithms
When it comes to planning (or control), evolutionary algo-
rithms are mostly used as follows: an evolutionary algorithm
is used in conjunction with an off-line simulator in order to
train a controller, for example as in [6]. The controller is
then used in the real problem. This approach is not just
used for control problems, but has been popular in fields
such as Artificial Life (e.g. [1]). This paper proposes a dif-
ferent approach. Evolution is used in the same manner as
MCTS uses roll-outs and the generative model (i.e. a sim-
ulator). Thus, an agent will evolve a plan in an imaginary
model for some milliseconds, acts on the (real) world by per-
forming the first action of its plan and then evolves a new
plan repeatedly (again in a simulation based manner) until
the game is over. This type of behaviour is called “Rolling
Horizon”, “Receding Horizon” or “Model Predictive” control
or planning. There have been some efforts in the past to
use such approaches (e.g. [17, 10], and this paper presents
two versions of a genetic algorithm in this setting. The term
Rolling Horizon comes from the fact that planning happens
until a limited look-ahead in the game (and thus we have to
keep on replanning at every time step). Contrary to MCTS,
we are currently unsure about the regret bounds of genetic
algorithms in the general case [15].
The Physical Travelling Salesman Problem (PTSP) is a
modification of the well known optimisation problem Trav-
elling Salesman Problem (or TSP).1In the PTSP, the player
controls a ship (or agent) that must visit a series of way-
points scattered around a continuous 2-D maze full of ob-
stacles. The score of the game is determined by the number
of waypoints visited and the time taken to do so. Figure 1
shows one of the maps of this game.
The PTSP is a real-time game. For every cycle, the ship
must supply an action to execute in no more than 40ms.
Additionally, there is a maximum time, depending on the
number of waypoints in the map, that cannot be overspent
when visiting the next waypoint. There are six different ac-
tions that can be provided to steer the ship. These actions
are discrete and are divided into two different inputs: ac-
celeration (which can be on or off ) and steering (which can
be left,right or straight). The action applied modifies the
position, velocity and orientation of the ship.
The PTSP featured in two competitions in 2012 which
1TSP consists of an agent who must visit all ncities once
via the shortest possible route, starting and ending at the
same location.
Figure 1: Sample PTSP Map.
took place at international IEEE conferences: the World
Congress On Computational Intelligence (WCCI) and the
Conference on Computational Intelligence and Games (CIG).
The interested reader can found more information about
the competition (game, rules and framework), as well as
background for TSP solvers, navigation algorithms and sim-
ilar real-time game competitions at [13]. The winner of
both competitions was a Monte Carlo Tree Search approach
based on macro-actions and a physics-based TSP route plan-
ner [14].
An obvious way to tackle the PTSP is by approaching the
two sub-problems it constitutes: the order of waypoints to
follow and the procedure to drive the ship to each of them.
In other words, it is necessary to tackle both short-term
and long-term planning tasks, whilst taking into account
the interdependencies between the two (i.e. the physics of
the game).
The focus of this paper is on the driving procedure. The
order of waypoints followed is the same for the different al-
gorithms presented in this study. This route is calculated
at the beginning of the game, during initialization, and it
remains available to the controller for the whole game. The
calculated route takes the physics of the game into account:
each given route has an associated cost assigned by a pro-
cedure that decomposes the route into straight-line obsta-
cle free segments. The time taken to complete each one of
these segments is estimated using the physics of the game
and keeping the speed at the end of the segments. A penal-
ization factor is applied at the end of each segment, related
to the angle between the current and next segment in the
route. This factor reduces the speed of the agent after each
segment, in order to approximate the effects of turns in the
trajectory of the ship. It has been determined [12] that those
routes involving fewer changes of direction, even when the
distance travelled is higher, usually provide better results.
4.1 Single versus Macro Actions
As stated in Section 3, there are six different actions that
can be applied at each time step. Each action must be de-
cided within the 40ms of time allowed for the controller to
choose a move. A good quality solution for a PTSP map
with 10 waypoints requires between 1000 (in the maps with
less obstacles) and 2000 (in more complex ones, as in the
one shown in Figure 1) time steps or action decisions. This
creates a search space of (1000,2000) (61000,62000 ), which
obviously makes it impossible to calculate the optimal route
within the time limitations given.
A reasonable reduction of this search space can be achieved
by limiting how far ahead in the game the algorithms can
look. As described before, the order of waypoints is calcu-
lated during the initialization process. The controller has in-
formation about how many and which waypoints have been
already visited, and it knows which ones are the next N
waypoints to be collected. If only the next 2 waypoints are
accounted, and according to the game length estimates, the
search space size is reduced to (62×100,62×200 ) (assuming
that the ship usually takes 100 to 200 actions per waypoint).
More reduction can be achieved by introducing the con-
cept of macro-actions. In this paper a macro-action is just
the repetition of the same action for Lconsecutive time
steps. The impact of changing a single action in a given
trajectory is not necessarily big, so high quality solutions
may still be found by not using such fine-grained precision.
For instance, choosing a macro-action length (defined as the
number of repetitions per macro-action) of 10, reduces the
search space size to (620,640 ). Different values of Lare used
during the experiments described in this paper in order to
show how this parameter affects the performance of each one
of the algorithms described.
Macro-actions have a second and important advantage:
they allow longer evaluations to execute before taking a
decision (through several game steps). In a single action
scenario, each algorithm has only 40ms to decide the next
action to execute. However, in the case of a macro-action
that is composed by Lsingle actions, each macro-action has
L×40ms to decide the move, as the previous macro-action
needs Ltime steps to be completed (with the exception of
the very first move, where there is no previous action). Al-
gorithm 1 defines a handler for macro-actions.
Algorithm 1 Algorithm to handle macro-actions.
1: function GetAction(GameState :gs)
2: if IsGameFirstAction(gs)then
3: actionT oRun DecideMacroAction(gs)
4: else
5: for i= 0 remainingActions do
6: gs.Advance(actionT oRun)
7: if remainingActions > 0then
8: if resetAlgorithm then
9: Algorithm.Reset(gs)
10: resetAlgorithm f alse
11: Algorithm.NextMove(gs)
12: else ⊲ remainingActions is 0
13: actionT oRun DecideMacroAction(gs)
14: remainingActions = (remainingActions 1)
15: return actionT oRun
17: function DecideMacroAction(GameState :gs)
18: actionT oRun Algorithm.NextMove(gs)
19: remainingActions L
20: resetAlgorithm true
21: return actionT oRun
The algorithm works as follows: the function GetAction,
called every game step, must return a (single) action to ex-
ecute in the current step. When the function is called for
the first time (line 3), or the number of remaining actions
is back to 0 (line 13), the function DecideMacroAction is
called. This function (lines 17 to 21) executes the decision
algorithm (which could be tree search, evolution or random
search), that is in charge of returning the move to make,
and sets a counter (remainingActions) to the length of the
macro-action (L). Before returning the action to run, it
indicates that the next cycle in the algorithm needs to be
In cases where the current step is not the first action of the
game, the state is advanced until the end of the macro-action
that is currently being executed (line 6). This way, the de-
cision algorithm can start the simulation from the moment
in time when the current macro-action will have finished.
The algorithm is then reset (if it was previously indicated
to do so by DecideMacroAction) and finally evaluates the
current state to take the next move (line 11).
Note that in this case the value returned by N extM ove is
ignored: the action to be executed is actionT oRun, decided
in a different time step. The objective of this call is to
keep building on the knowledge the algorithm is gathering
in order to decide the next macro-action (which effectively
happens in DecideM acroAction) while the ship makes the
move from the previous macro-action decision. Therefore,
Algorithm.NextM ove is called Ltimes, but only in the last
one is a real decision made.
It is important to note that the algorithms determine a
best solution of Nmacro-actions with Lsingle actions per
macro-action. However, only the first macro-action of the
best solution is finally handled and executed by the con-
troller. Once the best individual has been determined after
Literations, the first macro-action is used and the rest are
discarded, as the algorithm is reset at that point. The deci-
sion algorithm is in charge of keeping its internal state from
one call to the next, replying within the allowed time budget
and resetting itself when Algorithm.Reset is called.
4.2 Score function
Even the usage of macro-actions does not ensure that any
succession of Nactions reaches the next two waypoints in
the (pre-planned) route: it is unlikely that a random se-
quence of actions visits even one of them. In fact, the vast
majority of random trajectories will not visit any waypoints
at all. It is thus necessary to improve the granularity of the
rewards by defining a function that is able to distinguish the
quality of the different trajectories, by scoring the game in a
given state. This function is defined independently from the
algorithm that is going to use it, so the research is centred on
how the search space is explored by the different algorithms.
The score function takes into account the following pa-
rameters: distance and state (visited/unvisited) of the next
two waypoints in the route, time spent since the beginning of
the game and collisions in the current step. The final score
is the addition of four different values. The first one is given
by the distance points, which is defined as follows: being dw
the distance to the waypoint w, the reward for distance rdw
is set to 10000 dw. If the first waypoint is not visited, it is
set to rd1. However, if the first waypoint is visited, this is set
to rd1plus a reward value (10000). The other values for the
score function are given by the waypoints visited (number
of visited waypoints, out of the next two, multiplied by a re-
ward factor, 1000), time spent (set to 10000 minus the time
spent during the game) and collisions (100 if a collision is
happening in this step).
4.3 Genetic Algorithm
The two versions of the Genetic Algorithm (GA) employed
in this study are simple evolutionary algorithms, where each
individual encodes a group of macro-actions using integers.
Each one of the possible six actions is represented with a
number from 0 to 5. Note that each one of these genes
represents a macro-action: the action denoted by the gene
is repeated for Lsteps in the game where Lis defined in the
experimental setup.
All individuals in the population have a fixed length, as
defined in the experimental setup phase. As the algorithm
is meant to run online (i.e. while the game is actually being
played), a small population size of 10 individuals is used in
order to be able to evolve over a large number of generations.
Elitism is used to automatically promote the best 2 individ-
uals of each generation to the next. Evolution is stopped
when (40ms ǫ) is reached. ǫis introduced in order not to
use more time than allowed to provide an action.
Individuals are initialized uniformly at random such that
each gene can take any value between 0 and 5. This initial-
ization happens many times during each game: in the first
step of the game and every Lmoves, as described in Sec-
tion 4.1. Mutation is used with different rates: 0.2, 0.5 and
0.8. Each time a gene is mutated, only one of the two inputs
involved in the action is randomly chosen to be changed. If
acceleration is changed, the value is flipped from On to Off,
or vice versa. If steering is modified, left and right mutate
to straight, while straight can mutate to right or left at 50%
chance. This procedure was designed in order to transit
smoothly between actions. Tournaments of size 3 are used
to select individuals for uniform crossover.
Finally, the evaluation of an individual is performed by
the following mechanism: given an individual with Nmacro-
actions and Lsingle actions per macro-action, the game is
advanced, using the forward model mentioned in Section 3,
applying the N×Lactions defined in the individual. Then,
the resultant game state is evaluated with the score function
defined in Section 4.2.
Two different variants are employed in this study. The
first one (referred to as GAC), implements all features de-
scribed above. The second one (herein simply referred to as
GA) does not employ tournament selection nor crossover.
4.4 Monte Carlo Tree Search
MCTS is a simulation-based algorithm that builds a tree
in memory. Each node keeps track of statistics relating to
how often a move is played from a state (N(s, a)), how many
times each move is played from a state (N(s)) and the value
of the average reward (Q(s, a)). The algorithm presented
here uses the Upper Confidence Bound (UCB1; see Equa-
tion 1) equation at each decision node to balance between
exploration and exploitation. This balance is achieved by
setting the value of C, which provides the algorithm with
more or less exploration, at the expense of exploitation. A
commonly used value is 2, as it balances both facets of the
search when the rewards are normalized between 0 and 1.
a= arg max
aA(s)(Q(s, a) + Csln N(s)
N(s, a))(1)
Figure 2: MCTS algorithm steps.
The MCTS algorithm is executed in several steps [5]: at
the very beginning of the algorithm, the tree is composed
of only the root node, which is a representation of the cur-
rent state of the game. During the selection step, the tree
is navigated from the root until a maximum depth has been
reached. If a node has as many children as the available
number of actions from that position, UCB1 is used as tree
policy. Otherwise, a new node is added as a child of the cur-
rent one (expansion phase) and the simulation phase starts.
At this point, MCTS executes a Monte Carlo simulation
(or roll-out; default policy) from the expanded node until
the pre-defined depth, where the state of the game is eval-
uated. Finally, during the back-propagation step, the statis-
tics N(s), N(s, a) and Q(s, a) are updated for each node
traversed, using the reward obtained in the evaluation of
the state. These steps are executed in a loop until a termi-
nation criteria is met (in the case presented here, when the
time budget is consumed). Figure 2 summarizes the steps
of the algorithm.
In the problem presented in this paper, each of the actions
in the tree is a macro-action. This means that, when navi-
gating through the tree, Lsingle actions must be performed
to move from one state to the next, defined by the action
picked. In most games, the depth of the tree is limited by
the number of moves required to reach an end game state.
In the case of the PTSP, the end of the game is usually not
reached, so the maximum depth is determined by the num-
ber and length of the macro-actions, defined differently for
each experiment, bringing it closer to rolling horizon search
then normal MCTS.
The tree is preserved from one game step to the next (i.e.
during Lmoves) and is discarded every time the algorithm
is reset. When the controller asks for a move, the action
chosen is the one that corresponds to the child of the root
with the maximum expected value Q(s, a).
4.5 Random Search
Random Search (RS) is a simple algorithm that creates
random solutions of a specified length. During the allocated
time, new individuals are created uniformly at random and
evaluated, keeping the best solution found so far. As in
the other solvers, the best solution is forgotten every time
the algorithm is told to be reset. The motivation behind
including RS in this research is the good results obtained by
the GA with high mutation rates. Thus, it is important to
investigate how the results given by RS compare to those.
5.1 Experimental Setup
The experiments presented in this paper have been car-
ried out on 10 different maps, running 20 matches per maze.
Hence, each configuration has been executed during 200
games, in order to get reliable results.2The maps are those
distributed within the framework of the PTSP competition.3
Five different configurations are used for each algorithm,
varying the number of macro-actions (N) and the number of
single actions per macro-action (L). These configurations,
indicated by the pairs (N,L), are as follows:
The last four configurations share the property of foreseeing
the same distance into the future: N×L= 120. The only
exception is the first configuration, where L= 1, this is
considered as a special case as no macro-actions are used.
A key aspect for interpreting the results is to understand
how games are evaluated. If two single games are to be com-
pared, following the rules of PTSP, it is simple to determine
which game is better: the solution that visits more way-
points is the best. If both solutions visit the same number
of waypoints, the one that does so in the minimum number
of time steps wins.
However, when several matches are played in the same
map, determining the winner is not that simple. One ini-
tial approach is to calculate the averages of waypoints vis-
ited and time spent, following the same criteria used in the
single game scenario. The problem with this approach is
that the difference between two solutions, taken as the av-
erage of waypoints visited, can be quite small (less than one
waypoint), while the difference in time steps can be large.
For instance, imagine a scenario where two solvers (Aand
B) obtain an average of waypoints (wi) of wa= 9.51 and
wb= 9.47, and an average of time steps (ti) of ta= 1650
and tb= 1100. The traditional comparison would say that
Awins (because wa> wb), but actually Bis much faster
than A(tbta) with a very small waypoint difference.
Intuitively, Bseems to be doing a better job.
A possible solution to this problem is to calculate the ratio
ri=ti/withat approximates the time taken to visit a single
waypoint. The drawback to this approach is that it does not
provide a reliable comparison when one of the solvers visits
a small number of waypoints (or even 0). Therefore, the
following two features have been analysed:
Efficacy: number of waypoints visited, on average.
The higher this value, the better the solution provided.
Efficiency: ratio ri=ti/wi, but only for those matches
where wiequals the number of waypoints of the map.
The goal is to minimize this value, as it represents
faster drivers.
5.2 Results and analysis
5.2.1 40 milliseconds
Table 1 shows the number of waypoints visited on average,
including the standard error. The first notable aspect is
2All experiments were performed in a dedicated server Intel
Core i5 machine, 2.90GHz 6MB, and 4GB of memory.
that, out of the algorithms presented here, when no macro-
actions are provided (L= 1), only MCTS is able to solve the
problem up to a point. The other three algorithms cannot
complete the game, not getting more than one waypoint
in some rare cases. When the length of macro-action is
increased to 5, on average all algorithms achieve close to all
waypoints. Excellent efficacy is obtained where the macro-
action length is raised to 15, which seems to be the overall
best value for this parameter. In general, the GA (especially
with high mutation rates) seems to be the algorithm that
obtains the best efficacy.
Regarding efficiency, Figure 3 shows the ratio average with
standard error obtained with the GA (very similar results
are obtained with GAC). One of the first things to note is
that the mutation rate that obtains the best solutions is
0.5, for both evolutionary algorithms, especially when the
macro-action length is 10 or lower. It is interesting to see
how the distinct mutation rates make the algorithm behave
differently: it is clear that a higher mutation rate works
better with longer macro-actions, obtaining similar results to
0.5. On the other hand, although rate 0.5 is still the best, 0.2
provides better results than 0.8 for smaller values of L. This
is confirmed by a Mann-Whitney-Wilcoxon statistical test
(MW-test) run on this data: GA-0.5 is better than GA-0.8
for L= 5 and 10 (p-values of 0.006 and 0.041, respectively),
while GA-0.5 is faster than GA-0.2 for higher values of L
(p-values of 0.046 and 0.022 for L= 15,20).
Figure 4 shows the ratio average and standard error of
RS, MCTS, GA and GAC (both with mutation rate 0.5).
One initial result to note is that the 2012 PTSP winner
algorithm (MCTS) is no better than GA or GAC in any of
the macro-action lengths tested (with the exception, as men-
tioned earlier, of L= 1, which obtains a ratio of 168.37±7.7).
Actually, although overall there is no statistical difference,
GA and GAC get better results than MCTS in some of the
maps tested (see Section 5.3). It is also worthwhile mention-
ing that L= 15 provides the best results for all algorithms.
In fact, these results are compatible with previous studies:
Powley et al. [14] used L= 15 in their MCTS entry that
won the 2012 WCCI competition.
Finally, we observe an interesting phenomenon: while RS
is clearly worse than the other algorithms in lower macro
action-lengths (L= 5,10), the performance of all algorithms
converges to the same performance when the maximum value
of Lis reached.
Figure 3: Ratio (time/waypoints) results for GA.
A possible explanation for this effect is as follows: for the
maximum macro-action length (L= 20), the length of the
solution (or individual) that is being evaluated is N= 6, pro-
viding 663×105combinations. This evaluation performs
single actions during L×N= 120 time steps, but remember
that only the first macro-action is finally performed in the
actual game, discarding the rest of the individual actions (as
explained in Section 4.1). Clearly, not all the actions in the
sequence of 120 moves are equally important; actions chosen
at the beginning of this sequence produce a higher impact
in the overall trajectory followed by the ship. Therefore, it
is fair to assume that, for instance, the first half of the indi-
vidual (60 actions) is more relevant for the search than the
second. In the case of L= 20, this is encoded in the first 3
genes, which provides 63= 216 different possibilities. This
is a very reduced search space where even random search
can find a good solution (especially considering the amount
of evaluations performed per cycle, as shown in Section 5.4).
In the cases where L= 5 or L= 10, the same amount
of look ahead (60 single actions) is obtained with 12 and 6
genes, which provides a search space of 612 2×109, and
663×105respectively. In these more complex scenar-
ios, the search space is better explored by both GAs (and
MCTS) than by RS.
5.2.2 80 milliseconds
All the experiments described above were repeated with
the decision time doubled to 80ms. For the sake of space,
these results are not detailed here, but they are consistent
with the results obtained from spending 40ms per cycle.
All the algorithms perform slightly better when more time
is permitted, as would be expected. For instance, GA-0.5
obtains the best performance when L= 15, with a ratio
of 138.17 ±1.56, about three units faster than the results
achieved in 40ms. Again, the relationship between the dif-
ferent algorithms is maintained: all algorithms perform bet-
ter with L= 15 and their performances get closer as the size
of macro-action approaches 20.
5.3 Results by map
These results can also be analysed on a map-by-map basis.
Figure 5 represents the ratio achieved by each algorithm in
every map of this setup (remember lower is better). First of
all, it is interesting to see how the overall performance of the
algorithms changes depending on the map presented. For
Figure 4: Ratio (time/waypoints) results.
Average Waypoints Visited
LMCTS RS GA 0.2 GA 0.5 GA 0.8 GAC 0.2 GAC 0.5 GAC 0.8
1 7.66±0.25 0.03±0.01 0.03±0.01 0.05±0.01 0.01 ±0.01 0.03 ±0.01 0.02 ±0.01 0.01 ±0.01
5 9.62±0.11 9.76±0.08 9.87±0.06 9.89±0.06 9.86 ±0.06 9.91 ±0.05 9.9±0.05 9.69 ±0.09
10 9.62±0.12 9.95±0.03 9.96±0.03 10.0±0.0 10.0±0.0 9.96 ±0.04 9.91 ±0.06 9.96 ±0.03
15 10.0±0.0 10.0±0.0 10.0±0.0 10.0±0.0 10.0±0.0 10.0±0.0 10.0±0.0 10.0±0.0
20 9.97±0.03 9.95±0.04 10.0±0.0 10.0±0.0 10.0±0.0 9.97 ±0.02 9.98 ±0.02 9.92 ±0.04
Table 1: Waypoint visits with 40ms per time step.
Figure 5: Results per map, L= 15.
instance, map 9 (the one depicted in Figure 1) happens to
be especially complex. All algorithms obtain worse results
in this map than, for instance, in map 10, that contains
no obstacles except the boundaries of the map. Regarding
averages, GAC-0.5 obtains the best average of ratio per map
in 7 out of the 10 different maps, whilst MCTS leads the rest.
It is also possible to take the best results obtained in each
map and compare them with the ones obtained in the WCCI
PTSP competition. For instance, both GA and GAC obtain
better solutions than the winner of the WCCI competition
in 4 out of the 10 maps from the ones distributed with the
competition framework. Similarly, MCTS outperforms the
best results in two different maps.
The PTSP competition is quite strict regarding controllers
that overspend the 40 ms of time allowed per cycle, execut-
ing the action 0 (no acceleration, no steering) when this oc-
curs. In the experiments presented in this paper, however,
and for the sake of simplicity, this penalisation is ignored. In
other words, controllers do use 40ms, but should this time
limit be violated (which may happen if a single evaluation
takes more time than expected) the desired action is exe-
cuted anyway. In order to verify if the results obtained are
comparable with those from the competition, statistics re-
lating to the number of times the controllers overspent the
real-time limit were extracted. This phenomenon occurs in
only between 0.1% and 0.03% of the total game cycles in
a game, which suggests that, even if the results cannot be
fairly considered as new records for those maps, the perfor-
mance achieved by the algorithms presented in this paper is
significant and is comparable to some extent.
5.4 Number of evaluations
Another aspect worth mentioning is the number of evalu-
ations (or calls to the score function) performed every game
Figure 6: Evaluations per algorithm and L.
cycle. MCTS is able to perform more evaluations per cycle,
going from about 950 when L= 5 to 3000 if L= 20, while
the other algorithms range between 330 and 370. The main
reason for this being that MCTS stores intermediate states
of the game in the tree, and is able to reuse this information
during the simulations performed in a game cycle, whereas
the other three algorithms execute all the actions encoded
in each individual, from the first one to the last, producing
a more expensive and hence slower evaluation.
The difference between both GAs and RS depends on the
overhead of the algorithm (and not on the evaluation itself,
which takes the same time for all of them): as shown in
Figure 6, RS is the more efficient because of its simplicity,
followed by GA and then GAC, which is the slowest due
to the cost of using selection and crossover. M CT S is not
included in this picture for the sake of clarity.
It is interesting to see how both GAs obtain equivalent
(sometimes better) solutions than MCTS, the winner of the
2012 PTSP competition, even when they execute much fewer
evaluations per game cycle.
This paper compared the performance of several differ-
ent search techniques for the PTSP, a real-time game where
an action must be supplied every 40ms. The PTSP resem-
bles many modern games where the players need to move
providing actions within a few milliseconds, which allows to
extrapolate these conclusions to that type of games. The al-
gorithms considered were random search (RS), two different
evolutionary algorithms (GA and GAC) and Monte Carlo
Tree Search (MCTS). For this, a rolling horizon version for
the evolutionary algorithms is introduced. The search space
is by using macro-actions, where each one simply repeats a
low-level action a specified number of times, as has previ-
ously been shown to be successful [14]. The experiments
were performed on the maps distributed with the PTSP
competition enabling a direct comparison with the compe-
tition’s results.
MCTS was the only algorithm able to produce relatively
good solutions when no macro-actions were used. The fact
that this algorithm stores information in the tree allows for a
more efficient search when the search space is vast. However,
the best results were obtained using macro-actions, more
specifically where a macro-action length of 15 was used. In
this case, GAC and MCTS obtained better solutions on each
one of the maps, in some cases achieving better results than
those from the PTSP competition. The concept of macro
actions is in no way restricted to the PTSP and may equally
well be applied to the majority of real-time video games. Ex-
amples are high level commands in strategy games of first
person shooters, where the use of macro-actions instead of
low level orders enhances the performance of search algo-
rithms due to a reduction of the search space.
Another important contribution of this paper is under-
standing how different discretizations of the search space
affect the performance of the algorithms in this game. On
the one hand, if the granularity is fine (small macro-action
lengths, such as L= 5,10), the search algorithm can benefit
from artefacts such as the genetic operators. On the other
hand, if the granularity is coarse the fitness landscape is
rough, and the rolling horizon effect enables random search
to perform as well as the other algorithms. Finally, it is
worth mentioning that both GAs obtained very good so-
lutions while performing fewer evaluations than the other
techniques. This would suggest that the search for solutions
is more efficient in these two algorithms. This is signifi-
cant, suggesting that evolutionary algorithms offer an in-
teresting alternative to MCTS for general video game agent
controllers when applied in this on-line rolling horizon mode.
This work was supported by EPSRC grant EP/H048588/1.
[1] R. Beer and J. Gallagher. Evolving Dynamical Neural
Networks for Adaptive Behavior. Adaptive behavior,
1(1):91–122, 1992.
[2] C. Browne, E. Powley, D. Whitehouse, S. Lucas,
P. Cowling, P. Rohlfshagen, S. Tavener, D. Perez,
S. Samothrakis, and S. Colton. A Survey of Monte
Carlo Tree Search Methods. IEEE Transactions on
Computational Intelligence and AI in Games,
4:1:1–43, 2012.
[3] P.-A. Coquelin, R. Munos, et al. Bandit Algorithms
for Tree Search. In Uncertainty in Artificial
Intelligence, 2007.
[4] S. Edelkamp, P. Kissmann, D. Sulewski, and
H. Messerschmidt. Finding the Needle in the Haystack
with Heuristically Guided Swarm Tree Search. In
Multikonf. Wirtschaftsinform., pages 2295–2308,
Gottingen, Germany, 2010.
[5] S. Gelly, Y. Wang, R. Munos, and O. Teytaud.
Modification of UCT with Patterns in Monte-Carlo
Go. Technical report, Inst. Nat. Rech. Inform. Auto.
(INRIA), Paris, 2006.
[6] F. Gomez and R. Miikkulainen. Solving
Non-Markovian Control Tasks with Neuroevolution. In
International Joint Conference on Artificial
Intelligence, volume 16, pages 1356–1361.
[7] N. Ikehata and T. Ito. Monte Carlo Tree Search in Ms.
Pac-Man. In Proc. 15th Game Progrramming
Workshop, pages 1–8, Kanagawa, Japan, 2010.
[8] C.-S. Lee, M.-H. Wang, G. M. J.-B. Chaslot, J.-B.
Hoock, A. Rimmel, O. Teytaud, S.-R. Tsai, S.-C. Hsu,
and T.-P. Hong. The Computational Intelligence of
MoGo Revealed in Taiwan’s Computer Go
Tournaments. IEEE Transactions on Computational
Intelligence and AI in Games, 1(1):73–89, 2009.
[9] S. Matsumoto, N. Hirosue, K. Itonaga, K. Yokoo, and
H. Futahashi. Evaluation of Simulation Strategy on
Single-Player Monte-Carlo Tree Search and its
Discussion for a Practical Scheduling Problem. In
Proceedings of the International MultiConference of
Engineers and Computer Scientists, volume 3, pages
2086–2091, Hong Kong, 2010.
[10] J. Mercieca and S. Fabri. Particle Swarm
Optimization for Nonlinear Model Predictive Control.
In ADVCOMP 2011, The Fifth International
Conference on Advanced Engineering Computing and
Applications in Sciences, pages 88–93, 2011.
[11] D. Perez, P. Rohlfshagen, and S. Lucas. Monte-Carlo
Tree Search for the Physical Travelling Salesman
Problem. In Proceedings of EvoApplications, 2012.
[12] D. Perez, P. Rohlfshagen, and S. Lucas. Monte Carlo
Tree Search: Long Term versus Short Term Planning.
In Proceedings of the IEEE Conference on
Computational Intelligence and Games, 2012.
[13] D. Perez, P. Rohlfshagen, and S. Lucas. The Physical
Travelling Salesman Problem: WCCI 2012
Competition. In Proceedings of the IEEE Congress on
Evolutionary Computation, 2012.
[14] E. Powley, D. Whitehouse, and P. Cowling. Monte
Carlo Tree Search with Macro-actions and Heuristic
Route Planning for the Physical Travelling Salesman
Problem. In Proc. IEEE Conference of Computational
Intelligence in Games, pages 234–241, 2012.
[15] C. R. Reeves and J. E. Rowe. Genetic
Algorithms-Principles and Perspectives: a Guide to
GA Theory, volume 20. Kluwer Academic Pub, 2002.
[16] D. Robles and S. M. Lucas. A Simple Tree Search
Method for Playing Ms. Pac-Man. In Proceedings of
the IEEE Conference of Computational Intelligence in
Games, pages 249–255, Milan, Italy, 2009.
[17] S. Samothrakis and S. Lucas. Planning Using Online
Evolutionary Overfitting. In UK Workshop on
Computational Intelligence, pages 1–6. IEEE, 2010.
[18] S. Samothrakis, D. Robles, and S. M. Lucas. Fast
Approximate Max-n Monte-Carlo Tree Search for Ms
Pac-Man. IEEE Transactions on Computational
Intelligence and AI in Games, 3(2):142–154, 2011.
... Rolling Horizon Evolutionary Algorithm (RHEA) approaches game tree search as a planning problem and encodes sequences of actions as genomes (Perez et al., 2013). It uses standard evolutionary operators (recombination and mutation) to sparsely search for good sequences in the domain of possible agent actions up to some depth. ...
Full-text available
In this paper, we study AI approaches to successfully play a 2-4 players, full information, Bomberman variant published on the CodinGame platform. We compare the behavior of three search algorithms: Monte Carlo Tree Search, Rolling Horizon Evolution, and Beam Search. We present various enhancements leading to improve the agents' strength that concern search, opponent prediction, game state evaluation, and game engine encoding. Our top agent variant is based on a Beam Search with low-level bit-based state representation and evaluation function heavy relying on pruning unpromising states based on simulation-based estimation of survival. It reached the top one position among the 2,300 AI agents submitted on the CodinGame arena.
... Whereas most successful approaches for GVGAI games employ MCTS, it shall be noted that there are also other competitive approaches as the rolling horizon evolutionary algorithm (RHEA) [42] that evolve partial action sequences as a whole through an evolutionary optimization process. Furthermore, DL variants start to get used here as well [83]. ...
This paper reviews the field of Game AI, which not only deals with creating agents that can play a certain game, but also with areas as diverse as creating game content automatically, game analytics, or player modelling. While Game AI was for a long time not very well recognized by the larger scientific community, it has established itself as a research area for developing and testing the most advanced forms of AI algorithms and articles covering advances in mastering video games such as StarCraft 2 and Quake III appear in the most prestigious journals. Because of the growth of the field, a single review cannot cover it completely. Therefore, we put a focus on important recent developments, including that advances in Game AI are starting to be extended to areas outside of games, such as robotics or the synthesis of chemicals. In this article, we review the algorithms and methods that have paved the way for these breakthroughs, report on the other important areas of Game AI research, and also point out exciting directions for the future of Game AI.
... When used in a game-playing fashion (albeit uncommonly), these techniques are usually formatted to create a sequence of actions terminating at a rolling horizon. One such example is the Rolling Horizon Evolution Algorithm (RHEA) [16], which proved to be competitive against tree search agents in the General Video Game Artificial Intelligence Competition [10]. Justesen et al experimented with another technique known as online evolutionary planning in the game Hero Academy (Robot Entertainment 2012), which is a multi-action, turnbased, adversarial game [17]. ...
Full-text available
Search-based procedural content generation uses stochastic global optimization algorithms to search spaces of game content. However, it has been found that tree search can be competitive with evolution on certain optimization problems. We investigate the applicability of several tree search methods to map generation and compare them systematically with several optimization algorithms, including evolutionary algorithms. For purposes of comparison, we use a simplified map generation problem where only passable and impassable tiles exist, three different map representations, and a set of objectives that are representative of those commonly found in actual level generation problem. While the results suggest that evolutionary algorithms produce good maps faster, several tree search methods can perform very well given sufficient time, and there are interesting differences in the character of the generated maps depending on the algorithm chosen, even for the same representation and objective.
... It has also done very well on GVGAI. Another family of algorithms that has recently shown great promise on game planning tasks is rolling horizon evolution [36]. Here, evolutionary algorithms are used not for long-term learning, but for short-term action planning. ...
Conference Paper
If you are an artificial intelligence researcher, you should look to video games as ideal testbeds for the work you do. If you are a video game developer, you should look to AI for the technology that makes completely new types of games possible. This chapter lays out the case for both of these propositions. It asks the question "what can video games do for AI", and discusses how in particular general video game playing is the ideal testbed for artificial general intelligence research. It then asks the question "what can AI do for video games", and lays out a vision for what video games might look like if we had significantly more advanced AI at our disposal. The chapter is based on my keynote at IJCCI 2015, and is written in an attempt to be accessible to a broad audience.
This paper reviews the field of Game AI, which not only deals with creating agents that can play a certain game, but also with areas as diverse as creating game content automatically, game analytics, or player modelling. While Game AI was for a long time not very well recognized by the larger scientific community, it has established itself as a research area for developing and testing the most advanced forms of AI algorithms and articles covering advances in mastering video games such as StarCraft 2 and Quake III appear in the most prestigious journals. Because of the growth of the field, a single review cannot cover it completely. Therefore, we put a focus on important recent developments, including that advances in Game AI are starting to be extended to areas outside of games, such as robotics or the synthesis of chemicals. In this article, we review the algorithms and methods that have paved the way for these breakthroughs, report on the other important areas of Game AI research, and also point out exciting directions for the future of Game AI.
Full-text available
Algorithm UCB1 for multi-armed bandit problem has already been extended to Algorithm UCT (Upper bound Confidence for Tree) which works for minimax tree search. We have developed a Monte-Carlo Go program, MoGo, which is the first computer Go program using UCT. We explain our modification of UCT for Go application and also the intelligent random simulation with patterns which has improved significantly the performance of MoGo. UCT combined with pruning techniques for large Go board is discussed, as well as parallelization of UCT. MoGo is now a top level Go program on $9\times9$ and $13\times13$ Go boards.
Conference Paper
Full-text available
In this paper we investigate the use of Monte Carlo Tree Search (MCTS) on the Physical Travelling Salesman Problem (PTSP), a real-time game where the player navigates a ship across a map full of obstacles in order to visit a series of waypoints as quickly as possible. In particular, we assess the algorithm's ability to plan ahead and subsequently solve the two major constituents of the PTSP: the order of waypoints (long-term planning) and driving the ship (short-term planning). We show that MCTS can provide better results when these problems are treated separately: the optimal order of cities is found using Branch & Bound and the ship is navigated to collect the waypoints using MCTS. We also demonstrate that the physics of the PTSP game impose a challenge regarding the optimal order of cities and propose a solution that obtains better results than following the TSP route of minimum Euclidean distance.
Conference Paper
Full-text available
The significant success of MCTS in recent years, particularly in the game Go, has led to the application of MCTS to numerous other domains. In an ongoing effort to better understand the performance of MCTS in open-ended real-time video games, we apply MCTS to the Physical Travelling Salesman Problem (PTSP). We discuss different approaches to tailor MCTS to this particular problem domain and subsequently identify and attempt to overcome some of the apparent shortcomings. Results show that suitable heuristics can boost the performance of MCTS significantly in this domain. However, visualisations of the search indicate that MCTS is currently seeking solutions in a rather greedy manner, and coercing it to balance short term and long term constraints for the PTSP remains an open problem.
Conference Paper
Full-text available
Numerous competitions have emerged in recent years that allow researchers to evaluate their algorithms on a variety of real-time video games with different degrees of complexity. These competitions, which vary from classical arcade games like Ms Pac-Man to racing simulations (Torcs) and realtime strategy games (StarCraft), are essential to establish a uniform testbed that allows practitioners to refine their algorithms over time. In this paper we propose a new competition to be held for the first time at WCCI 2012: the Physical Travelling Salesman Problem is an open-ended single-player real-time game that removes some of the complexities evident in other video games while preserving some of the most fundamental challenges. This paper motivates and outlines the PTSP and discusses in detail the framework of the competition, including software interfaces, parameter settings, rules and details of submission.
Full-text available
Monte Carlo tree search (MCTS) is a recently proposed search method that combines the precision of tree search with the generality of random sampling. It has received considerable interest due to its spectacular success in the difficult problem of computer Go, but has also proved beneficial in a range of other domains. This paper is a survey of the literature to date, intended to provide a snapshot of the state of the art after the first five years of MCTS research. We outline the core algorithm's derivation, impart some structure on the many variations and enhancements that have been proposed, and summarize the results from the key game and nongame domains to which MCTS methods have been applied. A number of open research questions indicate that the field is ripe for future work.
We would like the behavior of the artificial agents that we construct to be as well-adapted to their environments as natural animals are to theirs. Unfortunately, designing controllers with these properties is a very difficult task. In this article, we demonstrate that continuous-time recurrent neural networks are a viable mechanism for adaptive agent control and that the genetic algorithm can be used to evolve effective neural controllers. A significant advantage of this approach is that one need specify only a measure of an agent's overall performance rather than the precise motor output trajectories by which it is achieved. By manipulating the performance evaluation, one can place selective pressure on the development of controllers with desired properties. Several novel controllers have been evolved, including a chemotaxis controller that switches between different strategies depending on environmental conditions, and a locomotion controller that takes advantage of sensory feedback if available but that can operate in its absence if necessary.
Conference Paper
We present a controller for the Physical Travelling Salesman Problem (PTSP), a path planning and steering problem in a simulated continuous real-time domain. Our approach is hierarchical, using domain-specific algorithms and heuristics to plan a coarse-grained route and Monte Carlo Tree Search (MCTS) to plan and steer along fine-grained paths. The MCTS component uses macro-actions to decrease the number of decisions to be made per unit of time and thus drastically reduce the size of the decision tree. Results from the 2012 WCCI PTSP Competition show that this approach significantly and consistently outperforms all other submitted AI controllers, and is competitive with strong human players. Our approach has potential applications to many other problems in movement planning and control, including video games.