Presentation is loading. Please wait.

Presentation is loading. Please wait.

1. Algorithms for Inverse Reinforcement Learning 2

Similar presentations


Presentation on theme: "1. Algorithms for Inverse Reinforcement Learning 2"— Presentation transcript:

1 1. Algorithms for Inverse Reinforcement Learning 2
1. Algorithms for Inverse Reinforcement Learning 2. Apprenticeship learning via Inverse Reinforcement Learning

2 Algorithms for Inverse Reinforcement Learning Andrew Ng and Stuart Russell

3 Motivation Given: (1) measurements of an agent's behavior over time, in a variety of circumstances, (2) if needed, measurements of the sensory inputs to that agent; (3) if available, a model of the environment. Determine: the reward function being optimized.

4 Why? Reason #1: Computational models for animal and human learning.
“In examining animal and human behavior we must consider the reward function as an unknown to be ascertained through empirical investigation.” Particularly true of multiattribute reward functions (e.g. Bee foraging: amount of nectar vs. flight time vs. risk from wind/predators)

5 Why? Reason #2: Agent construction.
“An agent designer [...] may only have a very rough idea of the reward function whose optimization would generate 'desirable' behavior.” e.g. “Driving well” Apprenticeship learning: Recovering expert's underlying reward function more “parsimonious” than learning expect's policy?

6 Possible applications in multi-agent systems
In multi-agent adversarial games, learning opponents’ reward functions that guild their actions to devise strategies against them. example In mechanism design, learning each agent’s reward function from histories to manipulate its actions. and more?

7 Inverse Reinforcement Learning (1) – MDP Recap
MDP is represented as a tuple (S, A, {Psa}, ,R) Note: R is bounded by Rmax Value function for policy : Q-function:

8 Inverse Reinforcement Learning (1) – MDP Recap
Bellman Equation: Bellman Optimality:

9 Inverse Reinforcement Learning (2) Finite State Space
Reward function solution set (a1 is optimal action)

10 Inverse Reinforcement Learning (2) Finite State Space
There are many solutions of R that satisfy the inequality (e.g. R = 0), which one might be the best solution? 1. Make deviation from as costly as possible: 2. Make reward function as simple as possible

11 Inverse Reinforcement Learning (2) Finite State Space
Va1 Va2 maximized Linear Programming Formulation: a1, a2, …, an R?

12 Inverse Reinforcement Learning (3) Large State Space
Linear approximation of reward function (in driving example, basis functions can be collision, stay on right lane,…etc) Let be value function of policy , when reward R = For R to make optimal

13 Inverse Reinforcement Learning (3) Large State Spaces
In an infinite or large number of state space, it is usually not possible to check all constraints: Choose a finite subset S0 from all states Linear Programming formulation, find αi that: x>=0, p(x)=x; otherwise p(x)=2x

14 Inverse Reinforcement Learning (4) IRL from Sample Trajectories
If is only accessible through a set of sampled trajectories (e.g. driving demo in 2nd paper) Assume we start from a dummy state s0,(whose next state distribution is according to D). In the case that reward trajectory state sequence (s0, s1, s2….):

15 Inverse Reinforcement Learning (4) IRL from Sample Trajectories
Assume we have some set of policies Linear Programming formulation The above optimization gives a new reward R, we then compute based on R, and add it to the set of policies reiterate

16 Discrete Gridworld Experiment
5x5 grid world Agent starts in bottom-left square. Reward of 1 in the upper-right square. Actions = N,W,S,E (30% chance of random)

17 Discrete Gridworld Results

18 Mountain Car Experiment #1
Car starts in valley, goal is at the top of hill Reward is -1 per “step” until goal is reached State = car's x-position & velocity (continuous!) Function approx. class: all linear combinations of 26 evenly spaced Gaussian-shaped basis functions

19 Mountain Car Experiment #2
Goal is in bottom of valley Car starts... not sure. Top of hill? Reward is 1 in the goal area, 0 elsewhere γ = 0.99 State = car's x-position & velocity (continuous!) Function approx. class: all linear combinations of 26 evenly spaced Gaussian-shaped basis functions

20 Mountain Car Results #1 #2

21 Continuous Gridworld Experiment
State space is now [0,1] x [0,1] continuous grid Actions: 0.2 movement in any direction + noise in x and y coordinates of [-0.1,0.1] Reward 1 in region [0.8,1] x [0.8,1], 0 elsewhere γ = 0.9 Function approx. class: all linear combinations of a 15x15 array of 2-D Gaussian-shaped basis functions m=5000 trajectories of 30 steps each per policy

22 Continuous Gridworld Results
3%-10% error when comparing fitted reward's optimal policy with the true optimal policy However, no significant difference in quality of policy (measured using true reward function)

23 Apprenticeship Learning via Inverse Reinforcement Learning
Pieter Abbeel & Andrew Y. Ng

24 Algorithm For t = 1,2,… Inverse RL step:
Estimate expert’s reward function R(s)= wT(s) such that under R(s) the expert performs better than all previously found policies {i}. RL step: Compute optimal policy t for the estimated reward w. Convey intuition about what the algorithm is doing. Find reward that makes teacher look much better; and then hypotesize that reward and find the policy. The intuition behind the algorithm, is that the set of policies generated gets closer (in some sense) to the expert at every iteration. RL black box. Inverse RL: explain the details now. Courtesy of Pieter Abbeel

25 Algorithm: IRL step Maximize , w:||w||2≤ 1 
s.t. Vw(E)  Vw(i) +  i=1,…,t-1  = margin of expert’s performance over the performance of previously found policies. Vw() = E [t t R(st)|] = E [t t wT(st)|] = wT E [t t (st)|] = wT () () = E [t t (st)|] are the “feature expectations” \mu(\pi) can be computed for any policy \pi based on the transition dynamics ‘’maybe mention – due to norm constraint -> QP, same as max margin for svm’s ; details  poster’’ Performance guarantees …  next slide Courtesy of Pieter Abbeel

26 Feature Expectation Closeness and Performance
If we can find a policy  such that ||(E) - ()||2  , then for any underlying reward R*(s) =w*T(s), we have that |Vw*(E) - Vw*()| = |w*T (E) - w*T ()|  ||w*||2 ||(E) - ()||2  . Let’s now look at how we could get performance guarantees, although the reward function is unknown. The key is that, if we can find a policy that is \epsilon close to the expert in feature expectations, then no matter what the underlying, unknown, reward function R* is, that policy will have performance \epsilon close to the expert’s performance as measure w.r.t. the unknown reward function. This is easily seen as follows. Going back to our algorithm, at every iteration it finds a policy closer to the expert in terms of feature expectations as follows. I will now illustrate this intuition graphically. In the inverse RL step it finds the direction in feature space for which the expert is mostly separated from the previously found policies. In the RL step, it finds a policy which is optimal w.r.t. this estimated reward, which brings it closer to the expert, again, in feature expectation space. Motivation for algorithm. --- get closer in \mu; Courtesy of Pieter Abbeel

27 IRL step as Support Vector Machine
maximum margin hyperplane seperating two sets of points () (E) |w*T (E) - w*T ()| = |Vw*(E) - Vw*()| = maximal difference between expert policy’s value function and 2nd to the optimal policy’s value function

28 2 (E) (2) w(3) (1) w(2) w(1) (0) 1 Uw() = wT()
Courtesy of Pieter Abbeel

29 Gridworld Experiment 128 x 128 grid world divided into 64 regions, each of size 16 x 16 (“macrocells”). A small number of macrocells have positive rewards. For each macrocell, there is one feature Φi(s) indicating whether that state s is in macrocell i Algorithm was also run on the subset of features Φi(s) that correspond to non-zero rewards.

30 Gridworld Results Distance to expert vs. # Iterations
Performance vs. # Trajectories

31 Car Driving Experiment
No explict reward function at all! Expert demonstrates proper policy via 2 min. of driving time on simulator (1200 data points). 5 different “driver types” tried. Features: which lane the car is in, distance to closest car in current lane. Algorithm run for 30 iterations, policy hand- picked. Movie Time! (Expert left, IRL right)

32 Demo-1 Nice

33 Demo-2 Right Lane Nasty

34 Car Driving Results


Download ppt "1. Algorithms for Inverse Reinforcement Learning 2"

Similar presentations


Ads by Google