Reinforcement Learning and NLP

Transcription

1 1 Reinforcement Learning and NLP Kapil Thadani RESEARCH

2 Outline 2 Model-free RL Markov decision processes (MDPs) Derivative-free optimization Policy gradients Variance reduction Value functions Actor-critic methods Policy gradients in NLP Non-differentiable metrics Latent structure

3 Reinforcement learning 3 Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child s? If this were then subjected to an appropriate course of education one would obtain the adult brain. Alan Turing Computing Machinery and Intelligence (1950)

4 Reinforcement learning 4 Sequential decision making Learn to model behavior over time Rewards may be stochastic and delayed Trade off exploration vs exploitation Generalization of supervised learning No full access to function to optimize Stateful environment, input affected by previous actions Nonstationarity for samples (no i.i.d. assumption)

5 Reinforcement learning 5 Agent Environment

6 Reinforcement learning 5 actions Agent Environment observations

7 Reinforcement learning 5 actions Agent Environment observations states rewards

8 Reinforcement learning 6 Model-free Policy-based: learn how to take actions in each state Value-based: learn the value of actions in each state Model-based Model environment to predict next states and rewards

9 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states A: set of actions P: transition probability distribution Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

10 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states s s A: set of actions P: transition probability distribution s Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

11 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states s s A: set of actions P: transition probability distribution s Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

12 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states s s A: set of actions P: transition probability distribution s Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

13 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states s s A: set of actions P: transition probability distribution s Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

14 Markov Decision Process (MDP) 7 Discrete-time stochastic control process defined by S, A, P, R, γ S: set of states s s A: set of actions P: transition probability distribution s Pss a = p(s t+1 = s s t = s, a t = a) R: reward function R a ss = E [r t s t = s, a t = a] γ [0, 1): discount factor

15 Markov Decision Process (MDP) 8 Goal: Take actions to maximize expected return over trajectories Trajectory τ: path through state space up to a horizon τ = s 0, a 0, r 0, s 1, a 1, r 1, s 2, a 2,... Episodic setting: agent acts until a terminal state is reached Return R t : cumulative future discounted rewards from s t R t = r t + γr t+1 + γ 2 r t+2 + = Assuming infinite horizon: if γ 1, R t is unbounded if 0 γ < 1, R t is well-defined and converges γ k r t+k k=0

16 Markov Decision Process (MDP) 9 Goal: Learn policy to maximize expected return over trajectories Policy π(s, a) represents action probabilities p(a s) in state s Deterministic policy: a t = π(s t) Stochastic policy: a t π(a s t) borrowing conditional prob notation Learn parameterized policy π θ with neural network weights θ Network architecture follows action space A: Discrete A: softmax layer to represent p(a s t) Continuous A: output µ and diagonal σ to sample a t N (µ, σ)

17 Derivative-free optimization 10 Objective: maximize E τ θ [R(τ)] Treat policy as a black box with parameters θ R d Iteratively update θ to make good returns more likely

18 11 Cross-entropy method Mannor et al (2003) The Cross Entropy method for Fast Policy Search Initialize µ 0 R d, σ 0 R d At each iteration i, draw L samples for θ l N (µ i, σi 2) Evaluate each trajectory τ l using parameters θ l Select the top ρ% of samples by R(τ l ) as the elite set Fit a new diagonal Gaussian to the elite set to obtain µ i+1, σ i+1 At convergence, return final µ + No gradients needed; only forward pass + Converges quickly + Remarkably effective on many problems, e.g., Tetris Needs lots of samples

19 12 Evolution strategies Salimans et al (2017) Evolution Strategies as a Scalable Alternative to Reinforcement Learning Initialize θ 0 R d At each iteration i, sample Gaussian noise ɛ 1,..., ɛ L N (0, I) Perturb θ i with each ɛ l to get θ l = θ i + σɛ l Evaluate each trajectory τ l using parameters θ l and update θ i+1 = θ i + η σl L R(τ l ) ɛ l l=1 + No gradients needed; only forward pass + Easy to parallelize with low communication overhead + Competitive on Atari, OpenAI benchmarks Requires 3-10x more data

20 Policy gradient 13 Learn to increase expected return by gradient ascent over θ θ i+1 = θ i + η θ E τ [R(τ)]

21 13 Policy gradient: reinforce Williams (1992) Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning Learn to increase expected return by gradient ascent over θ θ i+1 = θ i + η θ E τ [R(τ)] θ E τ [R(τ)] = τ = τ = τ R(τ) θ p(τ θ) R(τ) θ p(τ θ) p(τ θ) p(τ θ) R(τ) θ log p(τ θ) p(τ θ) = E τ [R(τ) θ log p(τ θ)] computed using sample averages 1 L R(τ l ) θ log p(τ l θ) L l=1

22 14 Policy gradient: reinforce Williams (1992) Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning Learn to increase expected return by gradient ascent over θ θ i+1 = θ i + η θ E τ [R(τ)] = θ i + η E τ [R(τ) θ log p(τ θ)] i.e., increase logprob of τ proportional to return R(τ) + Unbiased estimator of gradient + Valid even if R is discontinuous or unknown + Only need p(τ θ) to be differentiable High variance, particularly for long trajectories Assigns credit to whole trajectory rather than individual actions Large number of samples needed

23 Variance reduction: baseline If R(τ) 0 τ Estimator always modifies density θ i doesn t stabilize with fixed η 15 Subtract a baseline from the reward + Reduces variance θ E τ [R(τ)] = θ E τ [R(τ) b] + Estimator remains unbiased = E τ [(R(τ) b) θ log p(τ θ)] E τ b θ log p(τ θ) = b θ p(τ θ) = b θ 1 = 0 Using estimate of E [R(τ)] for b is a near-optimal choice i.e., increase logprob of τ proportional to how much return R(τ) is better than expected τ

24 Variance reduction 16 T 1 T 1 θ log p(τ θ) = θ log π θ (a t s t )Ps at ts t+1 = θ log π θ (a t s t ) t=0 t=0 + No need to model environment P sts t+1 (hence model-free) Rewards are distributed over trajectory [( T 1 ) T 1 ] θ E τ [R(τ)] = E τ r t θ log π θ (a t s t ) t=0 t=0 Instead, consider estimator for reward at timestep t [ ] t θ E τ [r t ] = E θ log π θ (a t s t ) r t t =0

25 Variance reduction 17 Sum over timesteps for full return [ T 1 θ E τ [R(τ)] = E τ Can add baseline for each state t=0 t=0 r t ] t θ log π θ (a t s t ) t =0 [ T 1 T 1 = E τ θ log π θ (a t s t ) [ T 1 θ E τ [R(τ)] = E τ θ log π θ (a t s t ) t=0 ( T 1 t =t r t ] )] r t b(s t ) t =t i.e., increase logprob of action a t proportional to how much future return R t (τ) = T 1 t =t r t is better than expected + Ignores prior rewards r 0,..., r t 1 when evaluating action a t at s t + Estimator remains unbiased as long as b doesn t depend on a t

26 Value functions 18 State-value function V π (s): Expected value of being in state s and following policy π V π (s) = E π [R t s t = s] State-action-value function Q π (s, a): Expected value of taking action a from s and then following π Q π (s) = E π [R t s t = s, a t = a] Advantage function A π (s, a): Expected value of taking action a from s instead of following π A π (s, a) = Q π (s, a) V π (s)

27 Value functions 19 [ T 1 T 1 θ E τ [R(τ)] = E τ θ log π θ (a t s t ) t=0 t =t r t ] + baseline term [ [ T 1 T 1 = E s0...a t θ log π θ (a t s t ) E rt...s T t=0 t =t r t Q π (s t, a t ) Substituting and reintroducing the baseline [ T 1 ] θ E τ [R(τ)] = E τ θ log π θ (a t s t ) (Q π (s t, a t ) b(s t )) t=0 ]] Defining b(s t ) = V π (s t ) is near-optimal (Greenmith et al, 2004) [ T 1 ] θ E τ [R(τ)] = E τ θ log π θ (a t s t )A π (s t, a t ) t=0

28 Value functions 20 [ T 1 ] θ E τ [R(τ)] = E τ θ log π θ (a t s t )A π (s t, a t ) t=0 Want E τ [A π ] = 0 to keep variance low i.e., positive advantage for good actions, negative for bad actions Don t know A π, can use an advantage estimator  such as Â(s t ) = r t + r t+1 + r t+2 + r t b(s t ) + Unbiased estimator High variance from single sample estimate Confounds the effect of actions a t, a t+1, a t+2,...

29 Variance reduction 21 Use discounted return when calculating advantage Â(s t ) = r t + γr t+1 + γ 2 r t+2 + γ 3 r t b(s t ) γ < 1 discounts the effect of actions that are far in the future To keep E τ [A π ] = 0, also use discounted return when fitting baseline to V π (s t ) + Emphasizes near-term rewards when evaluating actions + Lowers variance Biased estimator

30 Variance reduction 22 Use V π to estimate future rewards Â(s t ) = r t + γr t+1 + γ 2 r t+2 + γ 3 r t b(s t ) = r t + γr t+1 + γ 2 r t+2 + γ 3 V π (s t+3 ) b(s t ) = r t + γr t+1 + γ 2 V π (s t+2 ) b(s t ) = r t + γv π (s t+1 ) b(s t ) Use estimator of discounted V π for both future rewards and baseline Â(s t ) = r t + γ ˆV t (s t+1 ) ˆV t (s t ) + Explicit bias-variance tradeoff i.e., fewer reward terms lower variance of estimator but increase bias

31 Actor-critic methods 23 Â(s t ) = r t + γ ˆV (s t+1 ) ˆV (s t ) Actor: policy π θ (a t s t ) Critic: value function ˆV (s t ) Simplified algorithm: Evaluate policy π θi to collect samples s t, R t Fit ˆV by minimizing n ˆV (s n) R n 2 Update policy parameters using ˆV [ T 1 ( θ i+1 = θ i + η E τ θ log π θ (a t s t) r t + γ ˆV (s t+1) ˆV ) ] (s t) t=0

32 Resources 24 Reinforcement Learning: An Introduction (Sutton & Barto, 2017) OpenAI Spinning Up: code, environment, papers Berkeley RL course materials

33 Policy gradients in NLP 25 Optimization over non-differentiable metrics e.g., text generation metrics like BLEU, ROUGE, CIDEr etc Learned scores, e.g., neural teachers End-to-end training of sub-models for latent structure Sample from policy over latent variables Back-propagate loss to policy network

34 26 Non-differentiable metrics Rennie et al (2016) Self-critical Sequence Training for Image Captioning Optimize against CIDEr, BLEU, ROUGE, etc using reinforce Use score of greedy decoding of output words as a baseline i.e., encourage exploration of new words that improve rewards

35 26 Non-differentiable metrics Rennie et al (2016) Self-critical Sequence Training for Image Captioning Optimize against CIDEr, BLEU, ROUGE, etc using reinforce Use score of greedy decoding of output words as a baseline i.e., encourage exploration of new words that improve rewards

36 27 Non-differentiable metrics Bosselut et al (2018) Discourse-Aware Neural Rewards for Coherent Text Generation Train teacher network to score how well-ordered a text is Incorporate teacher scores as a reward in SCST (Rennie et al, 2016) to encourage generation of coherent text

37 28 Latent structure Andreas et al (2016) Learning to Compose Neural Networks for Question Answering Dynamically assemble a network for each question from modules Sample module layout from policy and backpropagate loss for answers to modules with reinforce

38 28 Latent structure Andreas et al (2016) Learning to Compose Neural Networks for Question Answering Dynamically assemble a network for each question from modules Sample module layout from policy and backpropagate loss for answers to modules with reinforce

39 29 Latent structure Lei et al (2016) Rationalizing Neural Predictions Define network to propose sequences of words from input text as rationales for aspect-based sentiment analysis Sample rationales during inference and convey squared error on task to rationale generator with reinforce

40 29 Latent structure Lei et al (2016) Rationalizing Neural Predictions Define network to propose sequences of words from input text as rationales for aspect-based sentiment analysis Sample rationales during inference and convey squared error on task to rationale generator with reinforce