Summary of part I: prediction and RL

Transcription

1 Summary of part I: prediction and RL Prediction is important for action selection The problem: prediction of future reward The algorithm: temporal difference learning Neural implementation: dopamine dependent learning in BG A precise computational model of learning allows one to look in the brain for hidden variables postulated by the model Precise (normative!) theory for generation of dopamine firing patterns Explains anticipatory dopaminergic responding, second order conditioning Compelling account for the role of dopamine in classical conditioning: prediction error acts as signal driving learning in prediction areas

2 prediction error hypothesis of dopamine measured firing rate at end of trial: δ t = r t - V t (just like R-W) V t = η t i=1 (1 η) t i r i model prediction error Bayer & Glimcher (2005)

3 Global plan Reinforcement learning I: prediction classical conditioning dopamine Reinforcement learning II: dynamic programming; action selection Pavlovian misbehaviour vigor Chapter 9 of Theoretical Neuroscience

4 Action Selection Evolutionary specification Immediate reinforcement: leg flexion Thorndike puzzle box pigeon; rat; human matching Delayed reinforcement: these tasks mazes chess Bandler; Blanchard

5 Immediate Reinforcement stochastic policy: based on action values: L m R m ; 5

6 use RW rule: Indirect Actor r L p L = R R 0.05; r = p 0.25 switch every 100 trials 6

7 Direct Actor L E ( m) = P[ L] r + P[ R] r R P[ L] L m = βp[ L] P[ R] P[ R] R m = βp[ L] P[ R] E ( m ) = β P [ L ] r P [ L ] r + P [ R ] r L m E( m) [ ]( L = β P L r E( m) ) L m E( m) β L m ( ( )) L L R ( L r E m ) ( ) if L is chosen L R L R a m m (1 ε )( m m ) + ε ( r E( m))( L R)

8 8 Direct Actor

9 Could we Tell? correlate past rewards, actions with present choice indirect actor (separate clocks): direct actor (single clock):

10 Matching: Concurrent VI-VI Lau, Glimcher, Corrado, Sugrue, Newsome

11 Matching income not return approximately exponential in r alternation choice kernel

12 Action at a (Temporal) Distance x=1 x=1 x=2 x=3 x=2 x=3 learning an appropriate action at x=1: depends on the actions at x=2 and x=3 gains no immediate feedback idea: use prediction as surrogate feedback 12

13 Action Selection L R start with policy: P[ L; x] = σ ( m ( x) m ( x)) x=1 x=2 x=3 x=1 x=2 x=3 evaluate it: V ( 1), V (2), V (3) improve it: x=1 x=2 x= thus choose R more frequently than L;C m * α δ

14 Policy δ > 0 if value is too pessimistic action is better than average v P x=1 x=2 x=3 14

15 actor/critic m 1 m 2 m 3 m n dopamine signals to both motivational & motor striatum appear, surprisingly the same suggestion: training both values & policies

16 Formally: Dynamic Programming

17 Variants: SARSA [ V x x u C] * t + ( t ) t = 1 t = ( actual (1, C) + ε r + Q(2, u ) Q(1, )) * Q ( 1, C) = E r + 1, Q( 1, C) Q C t Morris et al, 2006

18 Q * Variants: Q learning [ + V ( x ) x = 1 u C] * ( 1, C) = E r t+ 1 t, t t = ( r + max Q(2, u) Q(1, )) Q( 1, C) Q(1, C) + ε C t u Roesch et al, 2007

19 Summary prediction learning Bellman evaluation actor-critic asynchronous policy iteration indirect method (Q learning) asynchronous value iteration Q * V * [ ( ) 1] * t + V xt x = [ + V ( x ) x = 1 u C] * ( 1) E r 1 = + t ( 1, C) = E r t+ 1 t, t t =

20 Impulsivity & Hyperbolic Discounting humans (and animals) show impulsivity in: diets addiction spending, intertemporal conflict between short and long term choices often explained via hyperbolic discount functions alternative is Pavlovian imperative to an immediate reinforcer framing, trolley dilemmas, etc

21 Direct/Indirect Pathways Frank direct: D1: GO; learn from DA increase indirect: D2: nogo; learn from DA decrease hyperdirect (STN) delay actions given strongly attractive choices

22 Frank DARPP-32: D1 effect DRD2: D2 effect

23 Three Decision Makers tree search position evaluation situation memory

24 Multiple Systems in RL model-based RL build a forward model of the task, outcomes search in the forward model (online DP) optimal use of information computationally ruinous cached-based RL learn Q values, which summarize future worth computationally trivial bootstrap-based; so statistically inefficient learn both select according to uncertainty

25 Animal Canary OFC; dlpfc; dorsomedial striatum; BLA? dosolateral striatum, amygdala

26 Two Systems:

27 Behavioural Effects

28 Effects of Learning distributional value iteration (Bayesian Q learning) fixed additional uncertainty per step

29 One Outcome shallow tree implies goal-directed control wins

30 Human Canary... a b c if a c and c, then do more of a or b? MB: b MF: a (or even no effect)

31 Behaviour action values depend on both systems: Q tot ( x, u) = Q ( x, u) + βq ( x, u) MF MB expect that β will vary by subject (but be fixed)

32 Neural Prediction Errors (1 2) R ventral striatum (anatomical definition) note that MB RL does not use this prediction error training signal?

33 Neural Prediction Errors (1) right nucleus accumbens behaviour 1-2, not 1

34 Vigour Two components to choice: what: lever pressing direction to run meal to choose when/how fast/how vigorous free operant tasks real-valued DP 34

35 The model cost? how fast τ vigour cost C τ V LP unit cost (reward) C U P R U R NP S 0 Other τ 1 time S 1 S τ 2 time 2 choose (action,τ) = (LP,τ 1 ) Costs Rewards choose (action,τ) = (LP,τ 2 ) Costs Rewards goal 35

36 The model Goal: Choose actions and latencies to maximize the average rate of return (rewards minus costs per time) S 0 τ 1 time S 1 S τ 2 time 2 choose (action,τ) = (LP,τ 1 ) Costs Rewards choose (action,τ) = (LP,τ 2 ) Costs Rewards ARL 36

37 Average Reward RL Compute differential values of actions Differential value of taking action L with latency τ when in state x ρ = average rewards minus costs, per unit time Q L,τ (x) = Rewards Costs + C u C + v τ Future Returns V (x') τρ steady state behavior (not learning dynamics) 37 (Extension of Schwartz 1993)

38 Average Reward Cost/benefit Tradeoffs 1. Which action to take? Choose action with largest expected reward minus cost 2.How fast to perform it? slow less costly (vigour cost) slow delays (all) rewards net rate of rewards = cost of delay (opportunity cost of time) Choose rate that balances vigour and opportunity costs explains faster (irrelevant) actions under hunger, etc masochism 38

39 Optimal response rates probability 1 st Nose poke 0.4 rate per minute 30 Niv, Dayan, Joel, unpublished st NP LP seconds seconds since reinforcement Experimental data probability 1 st Nose poke seconds rate per minute seconds since reinforcement Model simulation 39

40 Optimal response rates Experimental data Model simulation % Responses on key A Pigeon A Pigeon B Perfect matching Herrnstein % Reinforcements on key A on lever A % Responses 50 0 Model Perfect matching More: 0 50 # responses % Reinforcements on lever A interval length amount of reward ratio vs. interval breaking point temporal structure etc. 40

41 Effects of motivation (in the model) RR25 Cv Q( x, u, τ ) = prr Cu + V ( x' ) τ R τ Q( x, u, τ ) Cv = R = 0 τ 2 opt = τ τ C R v opt mean latency low utility high utility energizing effect LP Other 41

42 Effects of motivation (in the model) RR25 directing effect response rate / minute response rate / minute seconds from reinforcement 1 U R 50% seconds from reinforcement low utility high utility mean latency 2 energizing effect LP Other 42

43 Relation to Dopamine Phasic dopamine firing = reward prediction error What about tonic dopamine? less more 43

44 Tonic dopamine = Average reward rate 1. explains pharmacological manipulations 2. dopamine control of vigour through BG pathways # LPs in 30 minutes Control DA depleted Aberman and Salamone 1999 # LPs in 30 minutes Control DA depleted eating time confound context/state dependence (motivation & drugs?) less switching=perseveration 44 NB. phasic signal RPE for choice/value learning Model simulation

45 Tonic dopamine hypothesis $ $ $ $ $ $ also explains effects of phasic dopamine on response times firing rate 45 reaction time Satoh and Kimura 2003 Ljungberg, Apicella and Schultz 1992

46 Sensory Decisions as Optimal Stopping consider listening to: decision: choose, or sample

47 Optimal Stopping equivalent of state u=1 is u 1 = n1 σ = C r 2.5 = 0.1 and states u=2,3 is u = ( n + n )

48 Transition Probabilities

49 Computational Neuromodulation dopamine phasic: prediction error for reward tonic: average reward (vigour) serotonin phasic: prediction error for punishment? acetylcholine: expected uncertainty? norepinephrine unexpected uncertainty; neural interrupt?

50 Conditioning 50 prediction: of important events control: in the light of those predictions Ethology optimality appropriateness Psychology classical/operant conditioning Computation dynamic progr. Kalman filtering Algorithm TD/delta rules simple weights Neurobiology neuromodulators; amygdala; OFC nucleus accumbens; dorsal striatum

51 Markov Decision Process class of stylized tasks with states, actions & rewards at each timestep t the world takes on state s t and delivers reward r t, and the agent chooses an action a t

52 Markov Decision Process World: You are in state 34. Your immediate reward is 3. You have 3 actions. Robot: I ll take action 2. World: You are in state 77. Your immediate reward is -7. You have 2 actions. Robot: I ll take action 1. World: You re in state 34 (again). Your immediate reward is 3. You have 3 actions.

53 Markov Decision Process Stochastic process defined by: reward function: r t ~ P(r t s t ) transition function: s t ~ P(s t+1 s t, a t )

54 Markov Decision Process Stochastic process defined by: reward function: r t ~ P(r t s t ) transition function: s t ~ P(s t+1 s t, a t ) Markov property future conditionally independent of past, given s t

55 The optimal policy Definition: a policy such that at every state, its expected value is better than (or equal to) that of all other policies Theorem: For every MDP there exists (at least) one deterministic optimal policy. by the way, why is the optimal policy just a mapping from states to actions? couldn t you earn more reward by choosing a different action depending on last 2 states?

56 Pavlovian & Instrumental Conditioning Pavlovian learning values and predictions using TD error Instrumental learning actions: by reinforcement (leg flexion) by (TD) critic (actually different forms: goal directed & habitual)

57 Pavlovian-Instrumental Interactions synergistic conditioned reinforcement Pavlovian-instrumental transfer Pavlovian cue predicts the instrumental outcome behavioural inhibition to avoid aversive outcomes neutral Pavlovian-instrumental transfer Pavlovian cue predicts outcome with same motivational valence opponent Pavlovian-instrumental transfer Pavlovian cue predicts opposite motivational valence negative automaintenance

58 -ve Automaintenance in Autoshaping simple choice task N: nogo gives reward r=1 G: go gives reward r=0 learn three quantities average value Q value for N Q value for G instrumental propensity is

59 -ve Automaintenance in Autoshaping Pavlovian action assert: Pavlovian impetus towards G is v(t) weight Pavlovian and instrumental advantages by ω competitive reliability of Pavlov new propensities new action choice

60 -ve Automaintenance in Autoshaping basic ve automaintenance effect (µ=5) lines are theoretical asymptotes equilibrium probabilities of action