G1. Factored MDPs
Malte Helmert and Gabriele R¨oger
Universit¨at Basel
Content of this Course
Planning
Classical
Foundations Logic Heuristics Constraints
Probabilistic
Explicit MDPs Factored MDPs
Content of this Course: Factored MDPs
Factored MDPs
Foundations Heuristic
Search Monte-Carlo
Methods
Factored MDPs
Factored MDPs
We would like to specify MDPs and SSPs with large state spaces.
In classical planning, we introducedplanning tasks to represent large transition systems compactly.
represent aspects of the world in terms of state variables states are avaluation of state variables
n (propositional) state variables induce 2n states
exponentially more compactthan “explicit” representation
Finite-Domain State Variables
Definition (Finite-Domain State Variable)
Afinite-domain state variableis a symbolvwith an associated domaindom(v), which is a finite non-empty set of values.
LetV be a finite set of finite-domain state variables.
Astate s overV is an assignments :V →S
v∈Vdom(v) such thats(v)∈dom(v) for all v ∈V.
AformulaoverV is a propositional logic formula whose atomic propositions are of the formv =d wherev ∈V andd ∈dom(v).
For simplicity, we only consider finite-domain state variables here.
Syntax of Operators
Definition (SSP and MDP Operators)
AnSSP operator o over a set of state variablesV has three components:
a preconditionpre(o), a logical formula over V
an effect eff(o) over V, defined on the following slides a costcost(o)∈R+0
AnMDP operator o over a set of state variablesV has three components:
a preconditionpre(o), a logical formula over V
an effect eff(o) over V, defined on the following slides a rewardreward(o) over V, defined on the following slides Whenever we just sayoperator(without SSP or MDP),
both kinds of operators are allowed.
Syntax of Effects
Definition (Effect)
Effectsover state variablesV are inductively defined as follows:
Ifv ∈V is a finite-domain state variable andd ∈dom(v), then v :=d is an effect (atomic effect).
Ife1, . . . ,en are effects, then (e1∧ · · · ∧en) is an effect (conjunctive effect).
The special case withn = 0 is theempty effect >.
Ife1, . . . ,en are effects andp1, . . . ,pn∈[0,1] such that Pn
i=1pi = 1, then(p1 :e1|. . .|pn:en) is an effect (probabilistic effect).
Note: To simplify definitions, conditional effects are omitted.
Effects: Intuition
Intuition for effects:
Atomic effectscan be understood as assignments that update the value of a state variable.
A conjunctive effecte = (e1∧ · · · ∧en) means that all subeffects e1, . . . , en take place simultaneously.
A probabilistic effecte = (p1 :e1|. . .|pn:en) means that exactly one subeffect ei ∈ {e1, . . . ,en}takes place with probability pi.
Semantics of Effects
Definition
Theeffect set[e] of an effecte is a set of pairshp,wi, wherep is a probability 0<p≤1 and w is a partial assignment. The effect set [e] is the set obtained recursively as
[v :=d] ={h1.0,{v 7→d}i}, [e∧e0] = ]
hp,wi∈[e],hp0,w0i∈[e0]
{hp·p0,w ∪w0i},
[p1:e1|. . .|pn:en] =
n
]
i=1
{hpi·p,wi | hp,wi ∈[ei]}.
whereU
is like S
but mergeshp,w0i andhp0,w0i tohp+p0,w0i.
Semantics of Operators
Definition (Applicable, Outcomes)
LetV be a set of finite-domain state variables.
Lets be a state over V, and let o be an operator overV. Operatoro isapplicable in s ifs |=pre(o).
Theoutcomesof applying an operatoro in s, written sJoK, are sJoK= ]
hp,wi∈[eff(o)]
{hp,sw0 i},
withsw0 (v) =d ifv =d ∈w andsw0 (v) =s(v) otherwise andU
is like S
but mergeshp,s0i andhp0,s0i to hp+p0,s0i.
Rewards
Definition (Reward)
Arewardover state variables V is inductively defined as follows:
c ∈Ris a reward
Ifχ is a propositional formula over V, [χ] is a reward
Ifr andr0 are rewards,r+r0,r−r0,r·r0 and rr0 are rewards Applying an MDP operatoro in s induces rewardreward(o)(s), i.e., the value of the arithmetic functionreward(o) where all occurrences ofv ∈V are replaced withs(v).
Probabilistic Planning Tasks
Probabilistic Planning Tasks
Definition (SSP and MDP Planning Task)
AnSSP planning task is a 4-tuple Π =hV,I,O, γiwhere V is a finite set of finite-domain state variables, I is a valuation over V called theinitial state, O is a finite set ofSSP operators overV, and γ is a formula overV called thegoal.
AnMDP planning task is a 4-tuple Π =hV,I,O,di where V is a finite set of finite-domain state variables, I is a valuation over V called theinitial state, O is a finite set ofMDP operators overV, and d ∈(0,1) is the discount factor.
Aprobabilistic planning task is an SSP or MDP planning task.
Mapping SSP Planning Tasks to SSPs
Definition (SSP Induced by an SSP Planning Task) The SSP planning task Π =hV,I,O, γi induces the SSPT =hS,A,c,T,s0,S?i, where
S is the set of all states over V, A is the set of operatorsO, c(o) =cost(o) for all o∈O, T(s,o,s0) =
(p ifo applicable in s andhp,s0i ∈sJoK 0 otherwise
s0=I, and
S? ={s ∈S |s |=γ}.
Mapping MDP Planning Tasks to MDPs
Definition (MDP Induced by an MDP Planning Task) The MDP planning task Π =hV,I,O,di induces the MDPT =hS,A,R,T,s0, γi, where
S is the set of all states over V, A is the set of operatorsO,
R(s,o) =reward(o)(s) for allo ∈O and s ∈S, T(s,o,s0) =
(p ifo applicable in s andhp,s0i ∈sJoK 0 otherwise
s0=I, and γ =d.
Complexity
Complexity of Probabilistic Planning
Definition (Policy Existence)
Policy existence (PolicyEx)is the following decision problem:
Given: SSP planning task Π
Question: Is there a proper policy for Π?
Membership in EXP
Theorem
PolicyEx∈EXP Proof.
The number of states in an SSP planning task is exponential in the number of variables. The induced SSP can be solved in time polynomial in|S| · |A|via linear programming and hence in time exponential in the input size.
EXP-completeness of Probabilistic Planning
Theorem
PolicyExisEXP-complete.
Proof Sketch.
Membership forPolicyEx: see previous slide.
Hardness is shown by Littman (1997) by reducing the EXP-complete gameG4 to PolicyEx.
Estimated Policy Evaluation
Large SSPs and MDPs
Before: optimal policies andexact state-values forsmall SSPs and MDPs.
Now: focus onlarge SSPs and MDPs Further algorithms not necessarily optimal (may generatesuboptimal policies)
Interleaved Planning & Execution
Number of reachable states of a policy usually exponentialin the number of state variables
For large SSPs and MDPs, policies cannot be provided explicitly.
Solution: (possibly approximate)compact representation of policy required to describe solution
⇒ not part of this lecture.
Alternative solution: interleave planning and execution
Interleaved Planning & Execution for SSPs
Plan-execute-monitor cyclefor SSP T: plan actiona for the current states execute a
observe new current states0 set s :=s0
repeat until s ∈S?
Interleaved Planning & Execution for MDPs
Plan-execute-monitor cyclefor MDP T: plan actiona for the current states execute a
observe new current states0 set s :=s0
repeat until discounted reward sufficiently small
Interleaved Planning & Execution in Practice
avoids loss of precisionthat often comes with compact description of policy
does not waste time with planning for states that are never reachedduring execution poor decisions can be avoided by
spending more time with planning before execution in SSPs, this can even mean that computed policy is not properand execution never reaches the goal in MDPs, it is not clear when the
discounted reward is sufficiently small
Estimated Policy Evaluation
The qualityof a policy is described by the state-value of the initial stateVπ(s0)
Quality of given policy π can be computed (viaLP or backward induction) or approximated arbitrarily closely (viaiterative policy evaluation) in small SSPs or MDPs Impossibleif planning and execution are interleaved as policy is incomplete
⇒Estimate quality of policy π byexecuting itn∈Ntimes
Executing a Policy
Definition (Run in SSP)
LetT be an SSP andπ be a proper policy for T. A sequence of transitions
ρπ =s0−p−−−−1:π(s0→) s1, . . . ,sn−1
pn:π(sn−1)
−−−−−−→sn is arunρπ ofπ ifsi+1∼siJπ(si)Kandsn∈S?. Thecostof run ρπ is cost(ρπ) =Pn−1
i=0 cost(π(si)).
A run in an SSP can easily be generated by executingπ froms0 until a states ∈S? is encountered.
Executing a Policy
Definition (Run in MDP)
LetT be an MDP andπ be a policy for T. A sequence of transitions
ρπ =s0
p1:π(s0)
−−−−−→s1, . . . ,sn−1
pn:π(sn−1)
−−−−−−→sn
is arunρπ ofπ ifsi+1∼siJπ(si)K.
Therewardof run ρπ is reward(ρπ) =Pn−1
i=0 γi·reward(si, π(si)).
To generate a run, a termination criterion (e.g., based on the change of the accumulated reward) must be specified.
Estimated Policy Evaluation
Definition (Estimated Policy Evaluation)
LetT be an SSP,π be a policy for T andhρ1π, . . . , ρnπi be a sequence of runs ofπ.
Theestimated qualityof π via estimated policy evaluationis V˜π := 1
n ·
n
X
i=1
cost(ρiπ).
Convergence of Estimated Policy Evaluation in SSPs
Theorem
LetT be an SSP,π be a policy for T andhρ1π, . . . , ρnπi be a sequence of runs ofπ.
ThenV˜π →Vπ(s0) for n→ ∞.
Proof.
Holds due to thestrong law of large numbers.
⇒V˜π is a good approximationof vπ(s0) ifn sufficiently large.
Summary
Summary
MDP and SSP planning tasks represent MDPs and SSPs compactly.
Policy existence in SSPs is EXP-complete.
Interleaving planning and execution avoids representation issues of (typically exponentially sized) policy.
Quality of such an incomplete policy can be estimatedby executing it a fixed number of times.
In SSPs, estimated policy evaluationconverges to the true quality of the policy.