Probabilistic Robotics Planning and Control Partially Observable Markov
Probabilistic Robotics Planning and Control: Partially Observable Markov Decision Processes SA-1
POMDPs n In POMDPs we apply the very same idea as in MDPs. n 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. n Let b be the belief of the agent about the state under consideration. n POMDPs compute a value function over belief space: 2
Problems n Each belief is a probability distribution, thus, each value in a POMDP is a function of an entire probability distribution. n This is problematic, since probability distributions are continuous. n Additionally, we have to deal with the huge complexity of belief spaces. n 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. 3
An Illustrative Example measurements state x 1 action u 3 state x 2 measurements actions u 1, u 2 payoff 4
The Parameters of the Example n n n The actions u 1 and u 2 are terminal actions. The action u 3 is a sensing action that potentially leads to a state transition. The horizon is finite and =1. 5
Payoff in POMDPs n n n 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: 6
Payoffs in Our Example (1) n n n If we are totally certain that we are in state x 1 and execute action u 1, we receive a reward of -100 If, on the other hand, we definitely know that we are in x 2 and execute u 1, the reward is +100. In between it is the linear combination of the extreme values weighted by the probabilities 7
Payoffs in Our Example (2) 8
The Resulting Policy for T=1 Given we have a finite POMDP with T=1, we would use V 1(b) to determine the optimal policy. n In our example, the optimal policy for T=1 is n n This is the upper thick graph in the diagram. 9
Piecewise Linearity, Convexity n The resulting value function V 1(b) is the maximum of the three functions at each point n It is piecewise linear and convex. 10
Pruning If we carefully consider V 1(b), we see that only the first two components contribute. n The third component can therefore safely be pruned away from V 1(b). n 11
Increasing the Time Horizon n Assume the robot can make an observation before deciding on an action. V 1(b) 12
Increasing the Time Horizon n Assume the robot can make an observation before deciding on an action. Suppose the robot perceives z 1 for which p(z 1 | x 1)=0. 7 and p(z 1| x 2)=0. 3. Given the observation z 1 we update the belief using Bayes rule. 13
Value Function V 1(b) b’(b|z 1) V 1(b|z 1) 14
Increasing the Time Horizon n n Assume the robot can make an observation before deciding on an action. Suppose the robot perceives z 1 for which p(z 1 | x 1)=0. 7 and p(z 1| x 2)=0. 3. Given the observation z 1 we update the belief using Bayes rule. Thus V 1(b | z 1) is given by 15
Expected Value after Measuring n Since we do not know in advance what the next measurement will be, we have to compute the expected belief 16
Expected Value after Measuring n Since we do not know in advance what the next measurement will be, we have to compute the expected belief 17
Resulting Value Function n The four possible combinations yield the following function which then can be simplified and pruned. 18
Value Function p(z 1) V 1(b|z 1) b’(b|z 1) p(z 2) V 2(b|z 2) 19
State Transitions (Prediction) n n When the agent selects u 3 its state potentially changes. When computing the value function, we have to take these potential state changes into account. 20
State Transitions (Prediction) 21
Resulting Value Function after executing u 3 n Taking the state transitions into account, we finally obtain. 22
Value Function after executing u 3 23
Value Function for T=2 n Taking into account that the agent can either directly perform u 1 or u 2 or first u 3 and then u 1 or u 2, we obtain (after pruning) 24
Graphical Representation of V 2(b) u 1 optimal u 2 optimal unclear outcome of measurement is important here 25
Deep Horizons and Pruning n n n 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 26
Deep Horizons and Pruning 27
28
Why Pruning is Essential n Each update introduces additional linear components to V. n Each measurement squares the number of linear components. n Thus, an un-pruned value function for T=20 includes more than 10547, 864 linear functions. n At T=30 we have 10561, 012, 337 linear functions. n The pruned value functions at T=20, in comparison, contains only 12 linear components. n The combinatorial explosion of linear components in the value function are the major reason why POMDPs are impractical for most applications. 29
POMDP Summary n n 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. 30
POMDP Approximations n Point-based value iteration n QMDPs n AMDPs 31
Point-based Value Iteration n Maintains a set of example beliefs n Only considers constraints that maximize value function for at least one of the examples 32
Point-based Value Iteration Value functions for T=30 Exact value function PBVI 33
Example Application 34
Example Application 35
QMDPs n QMDPs only consider state uncertainty in the first step n After that, the world becomes fully observable. 36
37
Augmented MDPs n Augmentation adds uncertainty component to state space, e. g. , Planning is performed by MDP in augmented state space n Transition, observation and payoff models have to be learned n 38
39
40
Coastal Navigation 41
Dimensionality Reduction on Beliefs 42
Monte Carlo POMDPs Represent beliefs by samples n Estimate value function on sample sets n Simulate control and observation transitions between beliefs n 43
Derivation of POMDPs Value Function Representation Piecewise linear and convex: 44
Value Iteration Backup in belief space: Belief update is a function: 45
Derivation of POMDPs Break into two components 46
Finite Measurement Space 47
Starting at Previous Belief *: constant **: linear function in params of belief space 48
Putting it Back in 49
Maximization over Actions 50
Getting max in Front of Sum 51
Final Result Individual constraints: 52
53
- Slides: 53