cse-573 reinforcement learning pomdps. planning what action next? perceptsactions environment static...
TRANSCRIPT
![Page 1: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/1.jpg)
CSE-573
Reinforcement LearningPOMDPs
![Page 2: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/2.jpg)
Planning
What action next?
Percepts Actions
Environment
Static vs. Dynamic
Fully vs.
Partially Observable
Perfectvs.
Noisy
Deterministic vs.
Stochastic
Discrete vs.
ContinuousOutcomes
Predictable vs. Unpredictable
![Page 3: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/3.jpg)
Classical Planning
What action next?
Percepts Actions
Environment
Static
Fully Observable
Perfect
Predictable
Discrete
Deterministic
![Page 4: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/4.jpg)
Stochastic Planning
What action next?
Percepts Actions
Environment
Static
Fully Observable
Perfect
Stochastic
Unpredictable
Discrete
![Page 5: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/5.jpg)
Markov Decision Process (MDP)
S: A set of states A: A set of actions Pr(s’|s,a): transition model R(s,a,s’): reward model s0: start state : discount factor
![Page 6: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/6.jpg)
Objective of a Fully Observable MDP
Find a policy : S → A
which maximizes expected discounted reward
given an infinite horizon
assuming full observability
![Page 7: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/7.jpg)
Define V*(s) {optimal value} as the maximum expected discounted reward from this state.
V* should satisfy the following equation:
V*(s) = maxaQ*(s,a) Solve using value/policy iteration, etc.
Bellman Equations for MDPs
Q*(s,a)
![Page 8: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/8.jpg)
Reinforcement Learning
Still have an MDP• Still looking for policy
New twist: don’t know Pr and/or R• i.e. don’t know which states are good• And what actions do
Must actually try actions and states out to learn
![Page 9: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/9.jpg)
Model based methods
Visit different states, perform different actions
Estimate Pr and R
Once model built, do planning using V.I. or other methods
Cons: require _huge_ amounts of data
![Page 10: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/10.jpg)
Model free methods
TD learning Directly learn Q*(s,a) values
sample = R(s,a,s’) + maxa’Qn(s’,a’) Nudge the old estimate towards the new
sample Qn+1(s,a) (1-)Qn(s,a) + [sample]
![Page 11: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/11.jpg)
Properties
Converges to optimal if• If you explore enough• If you make learning rate () small enough• But not decrease it too quickly
![Page 12: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/12.jpg)
Exploration vs. Exploitation
-greedy• Each time step flip a coin• With prob , action randomly• With prob 1- take the current greedy action
Lower over time to increase exploitation as more learning has happened
![Page 13: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/13.jpg)
Q-learning
Problems• Too many states to visit during learning• Q(s,a) is a BIG table
We want to generalize from small set of training examples
Solutions• Value function approximators• Policy approximators• Hierarchical Reinforcement Learning
![Page 14: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/14.jpg)
Task Hierarchy: MAXQ Decomposition [Dietterich’00]
Root
Take GiveNavigate(loc)
DeliverFetch
Extend-arm Extend-armGrab Release
MoveeMovewMovesMoven
Children of a task are
unordered
Children of a task are
unordered
![Page 15: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/15.jpg)
POMDPs
What action next?
Percepts Actions
Environment
Static
Partially Observable
Noisy
Stochastic
Unpredictable
Discrete
![Page 16: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/16.jpg)
Deterministic, fully observable
![Page 17: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/17.jpg)
Stochastic, Fully Observable
![Page 18: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/18.jpg)
Stochastic, Partially Observable
![Page 19: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/19.jpg)
21
POMDPs In POMDPs we apply the very same idea as in
MDPs.
Since the state is not observable, the agent has to make its decisions based on the belief state which is a posterior distribution over states.
Let b be the belief of the agent about the state under consideration.
POMDPs compute a value function over belief space:
γ
![Page 20: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/20.jpg)
22
Problems
Each belief is a probability distribution, thus, each value in a POMDP is a function of an entire probability distribution.
This is problematic, since probability distributions are continuous.
Additionally, we have to deal with the huge complexity of belief spaces.
For finite worlds with finite state, action, and measurement spaces and finite horizons, however, we can effectively represent the value functions by piecewise linear functions.
![Page 21: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/21.jpg)
23
An Illustrative Example
2x1x 3u8.0
2z1z
3u
2.0
8.02.0
7.0
3.0
3.0
7.0
measurements action u3 state x2
payoff
measurements
1u 2u 1u 2u
100 50100 100
actions u1, u2
payoff
state x1
1z
2z
![Page 22: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/22.jpg)
24
The Parameters of the Example The actions u1 and u2 are terminal actions.
The action u3 is a sensing action that potentially leads to a state transition.
The horizon is finite and =1.
![Page 23: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/23.jpg)
25
Payoff in POMDPs
In MDPs, the payoff (or return) depended on the state of the system.
In POMDPs, however, the true state is not exactly known.
Therefore, we compute the expected payoff by integrating over all states:
![Page 24: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/24.jpg)
26
Payoffs in Our Example (1)
If we are totally certain that we are in state x1 and execute action u1, we receive a reward of -100
If, on the other hand, we definitely know that we are in x2 and execute u1, the reward is +100.
In between it is the linear combination of the extreme values weighted by the probabilities
![Page 25: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/25.jpg)
27
Payoffs in Our Example (2)
![Page 26: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/26.jpg)
28
The Resulting Policy for T=1
Given we have a finite POMDP with T=1, we would use V1(b) to determine the optimal policy.
In our example, the optimal policy for T=1 is
This is the upper thick graph in the diagram.
![Page 27: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/27.jpg)
29
Piecewise Linearity, Convexity
The resulting value function V1(b) is the maximum of the three functions at each point
It is piecewise linear and convex.
![Page 28: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/28.jpg)
30
Pruning
If we carefully consider V1(b), we see that only the first two components contribute.
The third component can therefore safely be pruned away from V1(b).
![Page 29: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/29.jpg)
31
Increasing the Time Horizon Assume the robot can make an observation before
deciding on an action.
V1(b)
![Page 30: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/30.jpg)
32
Increasing the Time Horizon Assume the robot can make an observation before
deciding on an action. Suppose the robot perceives z1 for which
p(z1 | x1)=0.7 and p(z1| x2)=0.3. Given the observation z1 we update the belief using
Bayes rule.
3.04.0)1(3.07.0)(
)(
)1(3.0'
)(
7.0'
1111
1
12
1
11
pppzp
zp
pp
zp
pp
![Page 31: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/31.jpg)
33
Value Function
b’(b|z1)
V1(b)
V1(b|z1)
![Page 32: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/32.jpg)
34
Increasing the Time Horizon Assume the robot can make an observation before
deciding on an action. Suppose the robot perceives z1 for which
p(z1 | x1)=0.7 and p(z1| x2)=0.3. Given the observation z1 we update the belief using
Bayes rule. Thus V1(b | z1) is given by
![Page 33: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/33.jpg)
35
Expected Value after Measuring
Since we do not know in advance what the next measurement will be, we have to compute the expected belief
2
1111
2
1
111
2
1111
)|(
)(
)|()(
)|()()]|([)(
ii
i i
ii
iiiz
pxzpV
zp
pxzpVzp
zbVzpzbVEbV
![Page 34: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/34.jpg)
36
Expected Value after Measuring
Since we do not know in advance what the next measurement will be, we have to compute the expected belief
![Page 35: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/35.jpg)
37
Resulting Value Function The four possible combinations yield the
following function which then can be simplified and pruned.
![Page 36: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/36.jpg)
38
Value Function
b’(b|z1)
p(z1) V1(b|z1)
p(z2) V2(b|z2)
\bar{V}1(b)
![Page 37: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/37.jpg)
39
State Transitions (Prediction)
When the agent selects u3 its state potentially changes.
When computing the value function, we have to take these potential state changes into account.
![Page 38: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/38.jpg)
40
Resulting Value Function after executing u3
Taking the state transitions into account, we finally obtain.
![Page 39: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/39.jpg)
41
Value Function after executing u3
\bar{V}1(b)
\bar{V}1(b|u3)
![Page 40: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/40.jpg)
42
Value Function for T=2
Taking into account that the agent can either directly perform u1 or u2 or first u3 and then u1 or u2, we obtain (after pruning)
![Page 41: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/41.jpg)
43
Graphical Representation of V2(b)
u1 optimal u2 optimal
unclear
outcome of measuring is important here
![Page 42: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/42.jpg)
44
Deep Horizons and Pruning
We have now completed a full backup in belief space.
This process can be applied recursively. The value functions for T=10 and T=20 are
![Page 43: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/43.jpg)
45
Deep Horizons and Pruning
![Page 44: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/44.jpg)
46
![Page 45: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/45.jpg)
47
Why Pruning is Essential Each update introduces additional linear
components to V. Each measurement squares the number of
linear components. Thus, an unpruned value function for T=20
includes more than 10547,864 linear functions. At T=30 we have 10561,012,337 linear functions. The pruned value functions at T=20, in
comparison, contains only 12 linear components. The combinatorial explosion of linear components
in the value function are the major reason why POMDPs are impractical for most applications.
![Page 46: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/46.jpg)
48
POMDP Summary
POMDPs compute the optimal action in partially observable, stochastic domains.
For finite horizon problems, the resulting value functions are piecewise linear and convex.
In each iteration the number of linear constraints grows exponentially.
POMDPs so far have only been applied successfully to very small state spaces with small numbers of possible observations and actions.
![Page 47: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/47.jpg)
49
POMDP Approximations
Point-based value iteration
QMDPs
AMDPs
![Page 48: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/48.jpg)
50
Point-based Value Iteration
Maintains a set of example beliefs
Only considers constraints that maximize value function for at least one of the examples
![Page 49: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/49.jpg)
51
Point-based Value Iteration
Exact value function PBVI
Value functions for T=30
![Page 50: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/50.jpg)
52
QMDPs
QMDPs only consider state uncertainty in the first step
After that, the world becomes fully observable.
![Page 51: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/51.jpg)
53
N
jijjii xuxpxVuxruxQ
1
),|()(),(),(
N
jii
uuxQp
1
),(maxarg
![Page 52: CSE-573 Reinforcement Learning POMDPs. Planning What action next? PerceptsActions Environment Static vs. Dynamic Fully vs. Partially Observable Perfect](https://reader037.vdocuments.mx/reader037/viewer/2022103006/56649ece5503460f94bdb99e/html5/thumbnails/52.jpg)
54
Augmented MDPs
Augmentation adds uncertainty component to state space, e.g.
Planning is performed by MDP in augmented state space
Transition, observation and payoff models have to be learned
dxxbxbxH
xH
xbb b
b
x )(log)()(,)(
)(maxarg