This is the html version of the file http://zaphod.aml.sztaki.hu/papers/ecml06.pdf.
Google automatically generates html versions of documents as we crawl the web.
Bandit based Monte-Carlo Planning
 Page 1
Bandit based Monte-Carlo Planning
Levente Kocsis and Csaba Szepesvari
Computer and Automation Research Institute of the
Hungarian Academy of Sciences, Kende u. 13-17, 1111 Budapest, Hungary
kocsis@sztaki.hu
Abstract. For large state-space Markovian Decision Problems Monte-
Carlo planning is one of the few viable approaches to find near-optimal
solutions. In this paper we introduce a new algorithm, UCT, that ap-
plies bandit ideas to guide Monte-Carlo planning. In finite-horizon or
discounted MDPs the algorithm is shown to be consistent and finite
sample bounds are derived on the estimation error due to sampling. Ex-
perimental results show that in several domains, UCT is significantly
more efficient than its alternatives.
1 Introduction
Consider the problem of finding a near optimal action in large state-space Marko-
vian Decision Problems (MDPs) under the assumption a generative model of
the MDP is available. One of the few viable approaches is to carry out sampling
based lookahead search, as proposed by Kearns et al. [8], whose sparse lookahead
search procedure builds a tree with its nodes labelled by either states or state-
action pairs in an alternating manner, and the root corresponding to the initial
state from where planning is initiated. Each node labelled by a state is followed
in the tree by a fixed number of nodes associated with the actions available at
that state, whilst each corresponding state-action labelled node is followed by a
fixed number of state-labelled nodes sampled using the generative model of the
MDP. During sampling, the sampled rewards are stored with the edges connect-
ing state-action nodes and state nodes. The tree is built in a stage-wise manner,
from the root to the leafs. Its depth is fixed. The computation of the values of
the actions at the initial state happens from the leafs by propagating the values
up in the tree: The value of a state-action labelled node is computed based on
the average of the sum of the rewards along the edges originating at the node
and the values at the corresponding successor nodes, whilst the value of a state
node is computed by taking the maximum of the values of its children. Kearns
et al. showed that in order to find an action at the initial state whose value
is within the ϵ-vicinity of that of the best, for discounted MPDs with discount
factor 0 < γ < 1, K actions and uniformly bounded rewards, regardless of the
size of the state-space fixed size trees suffice [8]. In particular, the depth of the
tree is proportional to 1/(1−γ) log(1/(ϵ(1−γ))), whilst its width is proportional
to K/(ϵ(1 − γ)).
Although this result looks promising,
1
in practice, the amount of work needed
to compute just a single almost-optimal action at a given state can be over-
whelmingly large. In this paper we are interested in improving the performance
of this vanilla Monte-Carlo planning algorithm. In particular, we are interested in
Monte-Carlo planning algorithms with two important characteristics: (1) small
error probability if the algorithm is stopped prematurely, and (2) convergence
to the best action if enough time is given.
Besides MPDs, we are also interested in game-tree search. Over the years,
Monte-Carlo simulation based search algorithms have been used successfully in
many non-deterministic and imperfect information games, including backgam-
mon [13], poker [4] and Scrabble [11]. Recently, Monte-Carlo search proved to
be competitive in deterministic games with large branching factors, viz. in Go
[5]. For real-time strategy games, due to their enormous branching factors and
stochasticity, Monte-Carlo simulations seems to be one of the few feasible ap-
proaches for planning [7]. Intriguingly, Monte-Carlo search algorithms used by
today’s games programs use either uniform sampling of actions or some heuristic
biasing of the action selection probabilities that come with no guarantees.
The main idea of the algorithm proposed in this paper is to sample actions
selectively. In order to motivate our approach let us consider problems with a
large number of actions and assume that the lookahead is carried out at a fixed
depth D. If sampling can be restricted to say half of the actions at all stages
then the overall work reduction is (1/2)
D
. Hence, if one is able to identify a
large subset of the suboptimal actions early in the sampling procedure then
huge performance improvements can be expected.
By definition, an action is suboptimal for a given state, if its value is less than
the best of the action-values for the same state. Since action-values depend on the
values of successor states, the problem boils down to getting the estimation error
of the state-values for such states decay fast. In order to achieve this, an efficient
algorithm must balance between testing alternatives that look currently the best
so as to obtain precise estimates, and the exploration of currently suboptimal-
looking alternatives, so as to ensure that no good alternatives are missed because
of early estimation errors. Obviously, these criteria are contradictory and the
problem of finding the right balance is known as the the exploration-exploitation
dilemma. The most basic form of this dilemma shows up in multi-armed bandit
problems [1].
The main idea in this paper it to apply a particular bandit algorithm, UCB1
(UCB stands for Upper Confidence Bounds), for rollout-based Monte-Carlo plan-
ning. The new algorithm, called UCT (UCB applied to trees) described in Section
2 is called UCT. Theoretical results show that the new algorithm is consistent,
whilst experimental results (Section 3) for artificial game domains (P-games)
and the sailing domain (a specific MDP) studied earlier in a similar context by
others [10] indicate that UCT has a significant performance advantage over its
closest competitors.
1
In fact, as also noted by [8] the bound might be unimprovable, though this still
remains an open problem.
2 The UCT algorithm
2.1 Rollout-based planning
In this paper we consider Monte-Carlo planning algorithms that we call rollout-
based. As opposed to the algorithm described in the introduction (stage-wise
tree building), a rollout-based algorithm builds its lookahead tree by repeatedly
sampling episodes from the initial state. An episode is a sequence of state-action-
reward triplets that are obtained using the domains generative model. The tree
is built by adding the information gathered during an episode to it in an incre-
mental manner.
The reason that we consider rollout-based algorithms is that they allow us to
keep track of estimates of the actions’ values at the sampled states encountered
in earlier episodes. Hence, if some state is reencountered then the estimated
action-values can be used to bias the choice of what action to follow, potentially
speeding up the convergence of the value estimates. If the portion of states that
are encountered multiple times in the procedure is small then the performance
of rollout-based sampling degenerates to that of vanilla (non-selective) Monte-
Carlo planning. On the other hand, for domains where the set of successor states
concentrates to a few states only, rollout-based algorithms implementing selective
sampling might have an advantage over other methods.
The generic scheme of rollout-based Monte-Carlo planning is given in Fig-
ure 1. The algorithm iteratively generates episodes (line 3), and returns the
action with the highest average observed long-term reward (line 5).
2
In pro-
cedure UpdateValue the total reward q is used to adjust the estimated value
for the given state-action pair at the given depth, completed by increasing the
counter that stores the number of visits of the state-action pair at the given
depth. Episodes are generated by the search function that selects and effectu-
ates actions recursively until some terminal condition is satisfied. This can be
the reach of a terminal state, or episodes can be cut at a certain depth (line 8).
Alternatively, as suggested by Peret and Garcia [10] and motivated by itera-
tive deepening, the search can be implemented in phases where in each phase
the depth of search is increased. An approximate way to implement iterative
deepening, that we also follow in our experiments, is to stop the episodes with
probability that is inversely proportional to the number of visits to the state.
The effectiveness of the whole algorithm will crucially depend on how the
actions are selected in line 9. In vanilla Monte-Carlo planning (referred by MC
in the following) the actions are sampled uniformly. The main contribution of the
present paper is the introduction of a bandit-algorithm for the implementation
of the selective sampling of actions.
2.2 Stochastic bandit problems and UCB1
A bandit problem with K arms (actions) is defined by the sequence of random
payoffs X
it
, i = 1, . . ., K, t ≥ 1, where each i is the index of a gambling machine
2
The function bestMove is trivial, and is omitted due to the lack of space.
1: function MonteCarloPlanning(state)
2: repeat
3:
search(state, 0)
4: until Timeout
5: return bestAction(state,0)
6: function search(state, depth)
7: if Terminal(state) then return 0
8: if Leaf(state, d) then return Evaluate(state)
9: action := selectAction(state, depth)
10: (nextstate, reward) := simulateAction(state, action)
11: q := reward + γ search(nextstate, depth + 1)
12: UpdateValue(state, action, q, depth)
13: return q
Fig. 1. The pseudocode of a generic Monte-Carlo planning algorithm.
(the “arm” of a bandit). Successive plays of machine i yield the payoffs X
i1
,
X
i2
, . . .. For simplicity, we shall assume that X
it
lies in the interval [0, 1]. An
allocation policy is a mapping that selects the next arm to be played based
on the sequence of past selections and payoffs obtained. The expected regret
of an allocation policy A after n plays is defined by R
n
= max
i
E [∑
n
t=1
X
it
] −
E [∑
K
j=1
T
j
(n)
t=1
X
j,t
] , where I
t
∈ {1, . . ., K} is the index of the arm selected
at time t by policy A, and where T
i
(t) = ∑
t
s=1
I(I
s
= i) is the number of times
arm i was played up to time t (including t). Thus, the regret is the loss caused
by the policy not always playing the best machine. For a large class of payoff
distributions, there is no policy whose regret would grow slower than O(ln n)
[9]. For such payoff distributions, a policy is said to resolve the exploration-
exploitation tradeoff if its regret growth rate is within a constant factor of the
best possible regret rate.
Algorithm UCB1, whose finite-time regret is studied in details by [1] is
a simple, yet attractive algorithm that succeeds in resolving the exploration-
exploitation tradeoff. It keeps track the average rewards X
i,T
i
(t−1)
for all the
arms and chooses the arm with the best upper confidence bound:
I
t
= argmax
i∈{1,...,K}
{X
i,T
i
(t−1)
+ c
t−1,T
i
(t−1)
} ,
(1)
where c
t,s
is a bias sequence chosen to be
c
t,s
= √
2 ln t
s
.
(2)
The bias sequence is such that if X
it
were independently and identically dis-
tributed then the inequalities
P (X
is
≥ µ
i
+ c
t,s
) ≤ t
−4
,
(3)
P (X
is
≤ µ
i
− c
t,s
) ≤ t
−4
(4)
were satisfied. This follows from Hoeffding’s inequality. In our case, UCB1 is
used in the internal nodes to select the actions to be sampled next. Since for
any given node, the sampling probability of the actions at nodes below the node
(in the tree) is changing, the payoff sequences experienced will drift in time.
Hence, in UCT, the above expression for the bias terms c
t,s
needs to replaced by
a term that takes into account this drift of payoffs. One of our main results will
show despite this drift, bias terms of the form c
t,s
= 2C
p
ln t
s
with appropriate
constants C
p
can still be constructed for the payoff sequences experienced at any
of the internal nodes such that the above tail inequalities are still satisfied.
2.3 The proposed algorithm
In UCT the action selection problem as treated as a separate multi-armed bandit
for every (explored) internal node. The arms correspond to actions and the
payoffs to the cumulated (discounted) rewards of the paths originating at the
node.s In particular, in state s, at depth d, the action that maximises Q
t
(s, a, d)+
c
N
s,d
(t),N
s,a,d
(t)
is selected, where Q
t
(s, a, d) is the estimated value of action a in
state s at depth d and time t, N
s,d
(t) is the number of times state s has been
visited up to time t at depth d and N
s,a,d
(t) is the number of times action a was
selected when state s has been visited, up to time t at depth d.
3
2.4 Theoretical analysis
The analysis is broken down to first analysing UCB1 for non-stationary bandit
problems where the payoff sequences might drift, and then showing that the
payoff sequences experienced at the internal nodes of the tree satisfy the drift-
conditions (see below) and finally proving the consistency of the whole procedure.
The so-called drift-conditions that we make on the non-stationary payoff
sequences are as follows: For simplicity, we assume that 0 ≤ X
it
≤ 1. We assume
that the expected values of the averages X
in
=
1
n
n
t=1
X
it
converge. We let
µ
in
= E [X
in
] and µ
i
= lim
n→∞
µ
in
. Further, we define δ
in
by µ
in
= µ
i
in
and
assume that the tail inequalities (3),(4) are satisfied for c
t,s
= 2C
p
ln t
s
with an
appropriate constant C
p
> 0. Throughout the analysis of non-stationary bandit
problems we shall always assume without explicitly stating it that these drift-
conditions are satisfied for the payoff sequences.
For the sake of simplicity we assume that there exist a single optimal action.
4
Quantities related to this optimal arm shall be upper indexed by a star, e.g.,
µ
, T
(t), X
t
, etc. Due to the lack of space the proofs of most of the results are
omitted.
We let ∆
i
= µ
−µ
i
. We assume that C
p
is such that there exist an integer N
0
such that for s ≥ N
0
, c
ss
≥ 2|δ
is
| for any suboptimal arm i. Clearly, when UCT
is applied in a tree then at the leafs δ
is
= 0 and this condition is automatically
satisfied with N
0
= 1. For upper levels, we will argue by induction by showing an
upper bound δ
ts
for the lower levels that C
p
can be selected to make N
0
< +∞.
Our first result is a generalisation of Theorem 1 due to Auer et al. [1]. The
proof closely follows this earlier proof.
3
The algorithm has to be implemented such that division by zero is avoided.
4
The generalisation of the results to the case of multiple optimal arms follow easily.
Theorem 1 Consider UCB1 applied to a non-stationary problem. Let T
i
(n)
denote the number of plays of arm i. Then if i the index of a suboptimal arm,
n > K, then E [T
i
(n)] ≤
16C
2
p
ln n
(∆
i
/2)
2
+ 2N
0
+
π
2
3
.
At those internal nodes of the lookahead tree that are labelled by some state,
the state values are estimated by averaging all the (cumulative) payoffs for the
episodes starting from that node. These values are then used in calculating the
value of the action leading to the given state. Hence, the rate of convergence of
the bias of the estimated state-values will influence the rate of convergence of
values further up in the tree. The next result, building on Theorem 1, gives a
bound on this bias.
Theorem 2 Let X
n
= ∑
K
i=1
T
i
(n)
n
X
i,T
i
(n)
. Then
E [X
n
] − µ
≤ |δ
n
| + O (
K(C
2
p
ln n + N
0
)
n
) ,
(5)
UCB1 never stops exploring. This allows us to derive that the average rewards at
internal nodes concentrate quickly around their means. The following theorem
shows that the number of times an arm is pulled can actually be lower bounded
by the logarithm of the total number of trials:
Theorem 3 (Lower Bound) There exists some positive constant ρ such that
for all arms i and n, T
i
(n) ≥ ⌈ρ log(n)⌉.
Among the the drift-conditions that we made on the payoff process was that
the average payoffs concentrate around their mean quickly. The following result
shows that this property is kept intact and in particular, this result completes the
proof that if payoff processes further down in the tree satisfy the drift conditions
then payoff processes higher in the tree will satisfy the drift conditions, too:
Theorem 4 Fix δ > 0 and let ∆
n
= 9√2n ln(2/δ). The following bounds
hold true provided that n is sufficiently large: P (nX
n
≥ nE [X
n
] + ∆
n
) ≤ δ,
P (nX
n
≤ nE [X
n
] − ∆
n
) ≤ δ.
Finally, as we will be interested in the failure probability of the algorithm at the
root, we prove the following result:
Theorem 5 (Convergence of Failure Probability) LetI
t
= argmax
i
X
i,T
i
(t)
.
Then P(I
t
= i
) ≤ C (
1
t
)
ρ
2
min
i=i
∗ ∆i
36
2
with some constant C. In particular, it
holds that lim
t→∞
P(I
t
= i
) = 0.
Now follows our main result:
Theorem 6 Consider a finite-horizon MDP with rewards scaled to lie in the
[0, 1] interval. Let the horizon of the MDP be D, and the number of actions per
state be K. Consider algorithm UCT such that the bias terms of UCB1 are mul-
tiplied by D. Then the bias of the estimated expected payoff, X
n
, is O(log(n)/n).
Further, the failure probability at the root converges to zero at a polynomial rate
as the number of episodes grows to infinity.
Proof. (Sketch) The proof is done by induction on D. For D = 1 UCT just
corresponds to UCB1. Since the tail conditions are satisfied with C
p
= 1/√2 by
Hoeffding’s inequality, the result follows from Theorems 2 and 5.
Now, assume that the result holds for all trees of up to depth D − 1 and
consider a tree of depth D. First, divide all rewards by D, hence all the cumu-
lative rewards are kept in the interval [0, 1]. Consider the root node. The result
follows by Theorems 2 and 5 provided that we show that UCT generates a non-
stationary payoff sequence at the root satisfying the drift-conditions. Since by
our induction hypothesis this holds for all nodes at distance one from the root,
the proof is finished by observing that Theorem 2 and 4 do indeed ensure that
the drift conditions are satisfied. The particular rate of convergence of the bias
is obtained by some straightforward algebra.
By a simple argument, this result can be extended to discounted MDPs.
Instead of giving the formal result, we note that if some desired accuracy, ϵ
0
is fixed, similarly to [8] we may cut the search at the effective ϵ
0
-horizon to
derive the convergence of the action values at the initial state to the ϵ
0
-vicinity
of their true values. Then, similarly to [8], given some ϵ > 0, by choosing ϵ
0
small enough, we may actually let the procedure select an ϵ-optimal action by
sampling a sufficiently large number of episodes (the actual bound is similar to
that obtained in [8]).
3 Experiments
3.1 Experiments with random game trees
A P-game tree [12] is a minimax tree that is meant to model games where at
the end of the game the winner is decided by a global evaluation of the board
position where some counting method is employed (examples of such games
include Go, Amazons and Clobber). Accordingly, rewards are only associated
with transitions to terminal states. These rewards are computed by first assigning
values to moves (the moves are deterministic) and summing up the values along
the path to the terminal state.
5
If the sum is positive, the result is a win for
MAX, if it is negative the result is a win for MIN, whilst it is draw if the sum is 0.
In the experiments, for the moves of MAX the move value was chosen uniformly
from the interval [0, 127] and for MIN from the interval [−127, 0].
6
We have
performed experiments for measuring the convergence rate of the algorithm.
7
First, we compared the performance of four search algorithms: alpha-beta
(AB), plain Monte-Carlo planning (MC), Monte-Carlo planning with minimax
value update (MMMC), and the UCT algorithm. The failure rates of the four
algorithms are plotted as function of iterations in Figure 2. Figure 2, left cor-
responds to trees with branching factor (B) two and depth (D) twenty, and
5
Note that the move values are not available to the player during the game.
6
This is different from [12], where 1 and −1 was used only.
7
Note that for P-games UCT is modified to a negamax-style: In MIN nodes the
negative of estimated action-values is used in the action selection procedures.
1e-05
1e-04
0.001
0.01
0.1
1
1
10
100
1000
10000
100000
average error
iteration
B = 2, D = 20
UCT
AB
MC
MMMC
1e-05
1e-04
0.001
0.01
0.1
1
1
10
100
1000
10000
100000
1e+06
average error
iteration
B = 8, D = 8
UCT
AB
MC
MMMC
Fig. 2. Failure rate in P-games. The 95% confidence intervals are also shown for UCT.
Figure 2, right to trees with branching factor eight and depth eight. The failure
rate represents the frequency of choosing the incorrect move if stopped after a
number of iterations. For alpha-beta it is assumed that it would pick a move
randomly, if the search has not been completed within a number of leaf nodes.
8
Each data point is averaged over 200 trees, and 200 runs for each tree. We ob-
serve that for both tree shapes UCT is converging to the correct move (i.e. zero
failure rate) within a similar number of leaf nodes as alpha-beta does. Moreover,
if we accept some small failure rate, UCT may even be faster. As expected, MC
is converging to failure rate levels that are significant, and it is outperformed by
UCT even for smaller searches. We remark that failure rate for MMCS is higher
than for MC, although MMMC would eventually converge to the correct move
if run for enough iterations.
Second, we measured the convergence rate of UCT as a function of search
depth and branching factor. The required number of iterations to obtain failure
rate smaller than some fixed value is plotted in Figure 3. We observe that for
P-game trees UCT is converging to the correct move in order of B
D/2
number
of iterations (the curve is roughly parallel to B
D/2
on log-log scale), similarly to
alpha-beta. For higher failure rates, UCT seems to converge faster than o(B
D/2
).
Note that, as remarked in the discussion at the end of Section 2.4, due to
the faster convergence of values for deterministic problems, it is natural to de-
cay the bias sequence with distance from the root (depth). Accordingly, in the
experiments presented in this section the bias c
t,s
used in UCB was modified to
c
t,s
= (ln t/s)
(D+d)/(2D+d)
, where D is the estimated game length starting from
the node, and d is the depth of the node in the tree.
3.2 Experiments with MDPs
The sailing domain [10,14] is a finite state- and action-space stochastic shortest
path problem (SSP), where a sailboat has to find the shortest path between two
8
We also tested choosing the best looking action based on the incomplete searches.
It turns out, not unexpectedly, that this choice does not influence the results.
1
10
100
1000
10000
100000
1e+06
1e+07
4
6
8
10
12
14
16
18
20
iterations
depth
B = 2, D = 4-20
2^D
AB,err=0.000
UCT, err=0.000
UCT, err=0.001
UCT, err=0.01
UCT, err=0.1
2^(D/2)
10
100
1000
10000
100000
1e+06
1e+07
1e+08
8
7
6
5
4
3
2
iterations
branching factor
B = 2-8, D = 8
B^8
AB,err=0.000
UCT, err=0.000
UCT, err=0.001
UCT, err=0.01
UCT, err=0.1
B^(8/2)
Fig. 3. P-game experiment: Number of episodes as a function of failure probability.
points of a grid under fluctuating wind conditions. This domain is particularly
interesting as at present SSPs lie outside of the scope of our theoretical results.
The details of the problem are as follows: the sailboat’s position is represented
as a pair of coordinates on a grid of finite size. The controller has 7 actions
available in each state, giving the direction to a neighbouring grid position.
Each action has a cost in the range of 1 and 8.6, depending on the direction of
the action and the wind: The action whose direction is just the opposite of the
direction of the wind is forbidden. Following [10], in order to avoid issues related
to the choice of the evaluation function, we construct an evaluation function by
randomly perturbing the optimal evaluation function that is computed off-line
by value-iteration. The form of the perturbation isV(x) = (1 + ϵ(x))V
(x),
where x is a state, ϵ(x) is a uniform random variable drawn in [−0.1; 0.1] and
V
(x) is the optimal value function. The assignment of specific evaluation values
to states is fixed for a particular run. The performance of a stochastic policy is
evaluated by the error term Q
(s, a)−V
(s), where a is the action suggested by
the policy in state s and Q
gives the optimal value of action a in state s. The
error is averaged over a set of 1000 randomly chosen states.
Three planning algorithms are tested: UCT, ARTDP [3], and PG-ID [10]. For
UCT, the algorithm described in Section 2.1 is used. The episodes are stopped
with probability 1/N
s
(t), and the bias is multiplied (heuristically) by 10 (this
multiplier should be an upper bound on the total reward).
9
For ARTDP the
evaluation function is used for initialising the state-values. Since these values are
expected to be closer to the true optimal values, this can be expected to speed
up convergence. This was indeed observed in our experiments (not shown here).
Moreover, we found that Boltzmann-exploration gives the best performance with
ARTDP and thus it is used in this experiment (the ‘temperature’ parameter
is kept at a fixed value, tuned on small-size problems). For PG-ID the same
parameter setting is used as [10].
9
We have experimented with alternative stopping schemes. No major differences were
found in the performance of the algorithm for the different schemes. Hence these
results are not presented here.
1
10
100
1000
10000
100000
1e+06
1
10
100
1000
10000
samples
grid size
UCT
ARTDP
PG-ID
Fig. 4. Number of samples needed to achieve an error of size 0.1 in the sailing domain.
‘Problem size’ means the size of the grid underlying the state-space. The size of the
state-space is thus 24×‘problem size’, since the wind may blow from 8 directions, and
3 additional values (per state) give the ‘tack’.
Since, the investigated algorithms are building rather non-uniform search
trees, we compare them by the total number of samples used (this is equal to
the number of calls to the simulator). The required number of samples to obtain
error smaller than 0.1 for grid sizes varying from 2 × 2 to 40 × 40 is plotted in
Figure 4. We observe that UCT requires significantly less samples to achieve the
same error than ARTDP and PG-ID. At least on this domain, we conclude that
UCT scales better with the problem size than the other algorithms.
3.3 Related research
Besides the research already mentioned on Monte-Carlo search in games and the
work of [8], we believe that the closest to our work is the work of [10] who also
proposed to use rollout-based Monte-Carlo planning in undiscounted MDPs with
selective action sampling. They compared three strategies: uniform sampling (un-
controlled search), Boltzmann-exploration based search (the values of actions are
transformed into a probability distribution, i.e., samples better looking actions
are sampled more often) and a heuristic, interval-estimation based approach.
They observed that in the ‘sailing’ domain lookahead pathologies are present
when the search is uncontrolled. Experimentally, both the interval-estimation
and the Boltzmann-exploration based strategies were shown to avoid the looka-
head pathology and to improve upon the basic procedure by a large margin. We
note that Boltzmann-exploration is another widely used bandit strategy, known
under the name of “exponentially weighted average forecaster” in the on-line
prediction literature (e.g. [2]). Boltzmann-exploration as a bandit strategy is in-
ferior to UCB in stochastic environments (its regret grows with the square root of
the number of samples), but is preferable in adversary environments where UCB
does not have regret guarantees. We have also experimented with a Boltzmann-
exploration based strategy and found that in the case of our domains it performs
significantly weaker than the upper-confidence value based algorithm described
here.
Recently, Chang et al. also considered the problem of selective sampling in
finite horizon undiscounted MDPs [6]. However, since they considered domains
where there is little hope that the same states will be encountered multiple
times, their algorithm samples the tree in a depth-first, recursive manner: At
each node they sample (recursively) a sufficient number of samples to compute
a good approximation of the value of the node. The subroutine returns with an
approximate evaluation of the value of the node, but the returned values are
not stored (so when a node is revisited, no information is present about which
actions can be expected to perform better). Similar to our proposal, they suggest
to propagate the average values upwards in the tree and sampling is controlled
by upper-confidence bounds. They prove results similar to ours, though, due
to the independence of samples the analysis of their algorithm is significantly
easier. They also experimented with propagating the maximum of the values of
the children and a number of combinations. These combinations outperformed
propagating the maximum value. When states are not likely to be encountered
multiple times, our algorithm degrades to this algorithm. On the other hand,
when a significant portion of states (close to the initial state) can be expected
to be encountered multiple times then we can expect our algorithm to perform
significantly better.
4 Conclusions
In this article we introduced a new Monte-Carlo planning algorithm, UCT, that
applies the bandit algorithm UCB1 for guiding selective sampling of actions in
rollout-based planning. Theoretical results were presented showing that the new
algorithm is consistent in the sense that the probability of selecting the optimal
action can be made to converge to 1 as the number of samples grows to infinity.
The performance of UCT was tested experimentally in two synthetic do-
mains, viz. random (P-game) trees and in a stochastic shortest path problem
(sailing). In the P-game experiments we have found that the empirically that
the convergence rates of UCT is of order B
D/2
, same as for alpha-beta search
for the trees investigated. In the sailing domain we observed that UCT requires
significantly less samples to achieve the same error level than ARTDP or PG-
ID, which in turn allowed UCT to solve much larger problems than what was
possible with the other two algorithms.
Future theoretical work should include analysing UCT in stochastic shortest
path problems and taking into account the effect of randomised terminating
condition in the analysis. We also plan to use UCT as the core search procedure
of some real-world game programs.
5 Acknowledgements
We would like to acknowledge support for this project from the Hungarian Na-
tional Science Foundation (OTKA), Grant No. T047193 (Cs. Szepesvari) and
from the Hungarian Academy of Sciences (Cs. Szepesvari, Bolyai Fellowship).
References
1. P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite time analysis of the multiarmed
bandit problem. Machine Learning, 47(2-3):235–256, 2002.
2. P. Auer, N. Cesa-Bianchi, Y. Freund, and R.E. Schapire. The nonstochastic mul-
tiarmed bandit problem. SIAM Journal on Computing, 32:48–77, 2002.
3. A.G. Barto, S.J. Bradtke, and S.P. Singh. Real-time learning and control using
asynchronous dynamic programming. Technical report 91-57, Computer Science
Department, University of Massachusetts, 1991.
4. D. Billings, A. Davidson, J. Schaeffer, and D. Szafron. The challenge of poker.
Artificial Intelligence, 134:201–240, 2002.
5. B. Bouzy and B. Helmstetter. Monte Carlo Go developments. In H.J. van den
Herik, H. Iida, and E.A. Heinz, editors, Advances in Computer Games 10, pages
159–174, 2004.
6. H.S. Chang, M. Fu, J. Hu, and S.I. Marcus. An adaptive sampling algorithm for
solving Markov decision processes. Operations Research, 53(1):126–139, 2005.
7. M. Chung, M. Buro, and J. Schaeffer. Monte Carlo planning in RTS games. In
CIG 2005, Colchester, UK, 2005.
8. M. Kearns, Y. Mansour, and A.Y. Ng. A sparse sampling algorithm for near-
optimal planning in large Markovian decisi on processes. In Proceedings of IJ-
CAI’99, pages 1324–1331, 1999.
9. T.L. Lai and H. Robbins. Asymptotically efficient adaptive allocation rules. Ad-
vances in Applied Mathematics, 6:4–22, 1985.
10. L. Peret and F. Garcia. On-line search for solving Markov decision processes via
heuristic sampling. In R.L. de Mantaras and L. Saitta, editors, ECAI, pages 530–
534, 2004.
11. B. Sheppard. World-championship-caliber Scrabble. Artificial Intelligence, 134(1–
2):241–275, 2002.
12. S.J.J. Smith and D.S. Nau. An analysis of forward pruning. In AAAI, pages
1386–1391, 1994.
13. G. Tesauro and G.R. Galperin. On-line policy improvement using Monte-Carlo
search. In M.C. Mozer, M.I. Jordan, and T. Petsche, editors, NIPS 9, pages 1068–
1074, 1997.
14. R. Vanderbei. Optimal sailing strategies, statistics and operations research pro-
gram. University of Princeton, http://www.sor.princeton.edu/˜rvdb/sail/sail.html.,
1996.