1. Optimal Decision-Making in Humans & Animals Angela Yu March 05, 2009

2. Introduction A simple decision-making (DM) task binary decision rate of information flow constant over time task difficulty easily manipulated

3. Random-Dot Coherent Motion Paradigm 30% Coherence5% Coherence

4. Neurobiology Saccade generation system LIP response vs. coherence (Roitman & Shalden, 2002; Gold & Shadlen, 2004)

5. Behavioral Data Accuracy vs. Coherence vs. Coherence (Roitman & Shadlen, 2002)

6. A Need for Theoretical Framework for DM Speed Accuracyvs. Faster responses save time, but more errors What constitutes an optimal policy? Are humans/animals optimal? How is it neurally implemented?

7. Review of Bayesian decision theory: examples Quantifying temporal cost Optimal policy for 2-alternative forced choice – optimal policy (SPRT) – relation to data Deciding under a deadline Outline

8. Bayesian Decision Theory: the Idea s x hidden variable observation ? loss function

9. mean of p(s|x) is optimal Bayesian Decision Theory: Other Examples median of p(s|x) is optimal Ex. 2: loss function is absolute error: Ex. 1: loss function is mean square error:

10. Sequential Data  Repeated Decisions 010101 wait +: more info -: more time (Bayes’ Rule)

11. Loss function penalizes both error and delay: Optimal Policy (Bellman equation): Loss Function and Optimal Policy Stopping cost: Expected loss: Continuation cost: After observing inputs x 1, …, x t, continue (observe x t+1 ) only if the continuation cost < stopping cost

12. An Intractable Solution in Practice Stopping Cost Too hard!

13. b a Optimal Decision Policy: SPRT Wald & Wolfowitz (1948): Continuation region is a fixed interval (a, b) on the unit interval: At time t, continue if b > q t > a, stop and choose = 1 if q t > b, report = 0 if q t < a. example trajectories

14. Intuition for Optimality of SPRT Cost qtqt 01 c ab Q s (0) = w q t Q s (1) = v (1 - q t ) Q c is concave decide 0decide 1continue

15. Sequential Probability Ratio Test Log posterior ratio (monotonic to q t ): b’ a’ 0 rtrt Additive updates (random walk):

16. Relating Back to Data AccuracyMean RT (Data from Roitman & Shadlen, 2002; analysis from Ditterich, 2007) BehaviorBiology LIP response vs. coherence

17. Mid-way Summary Optimal binary DM = evidence accumulation to thresholds Perceptual DM behavior similar to that expected from SPRT Neural activities in LIP  Bayesian evidence integration SPRT a formal link between biology and behavior in DM

18. Review of Bayesian decision theory: examples Quantifying temporal cost Optimal policy for 2-alternative forced choice – optimal policy (SPRT) – relation to data Deciding under a deadline Outline

19. Caveat: Model Fit Imperfect Mean RTRT Distributions (Data from Roitman & Shadlen, 2002; analysis from Ditterich, 2007) SPRT predicts same RT distribution for correct & error trials Real error trials are slower

20. b a SPRT: Same RT for Correct and Errors

21. Fix 1: Variable Drift Rate (Data from Roitman & Shadlen, 2002; analysis from Ditterich, 2007) (idea from Ratcliff & Rouder, 1998) AccuracyMean RTRT Distributions

22. Fix 2: Increasing Drift Rate (Data from Roitman & Shadlen, 2002; analysis from Ditterich, 2007) AccuracyMean RTRT Distributions

23. A Principled Approach Trials aborted w/o response or reward if monkey breaks fixation (Data from Roitman & Shadlen, 2002) (Analysis from Ditterich, 2007) More “urgency” over time as risk of aborting trial increases Can we model this in a principled way?

24. Optimal Policy is a pair of of monotonically decaying thresholds Probability 1 0 Generalized SPRT Classic SPRT Time Intuition for decay With limited time, there is limited opportunity to observe enough evidence to bias the evidence in the opposite direction Optimal Decision Making Under a Deadline (Frazier & Yu, 2007) Loss function depends on error, delay, and whether deadline exceed

25. Threshold Decay  Fast RT & Slower Errors Probability 1 0 Generalized SPRT Classic SPRT Time (Frazier & Yu, 2007) Earlier deadline  greater urgency Time Probability

26. Timing Uncertainty (Frazier & Yu, 2007) Timing uncertainty  more conservative policy (lower threhsolds) Time Probability This leads to a set of testable predictions

27. Generalizing the Cost Function Loss function increases faster than linear in time f(  ) is convex and increasing. Nonlinear trade-off between accuracy and time (e.g. maximizing reward rate)

28. Summary Bayesian decision theory as a formal framework for DM Inadequacies of SPRT to account for behavior Inherent time pressure  decaying thresholds Neural implementation of decaying thresholds?

29. Perceptual DM + Saccade Planning Task: find & identify target sensory processing cost (time, accuracy) learning/memory attention Current/Future Project BehaviorTheoryNeurobiology Human Rats Bayesian modeling Stochastic control theory fMRI Recording

31. Bayesian Decision Theory Decision policy Loss function: Expected loss: A policy  : x  d is a mapping from input x into decision d, where x is a sampled from the distribution p(x|s). The loss function l(d(x); s) depends on the decision d(x) and the hidden variable s. The expected loss associated with a policy  is averaged over the prior distribution p(s) and the likelihood p(x|s): The optimal policy minimizes this expected loss.

32. Examples Ex. 1: loss function is mean square error For each observation x, we need to minimize: Differentiating and setting to 0, we find loss is minimized if: Optimal policy is to report the mean of posterior p(s|x).

33. Examples Ex. 2: loss function is absolute error: Optimal policy is to report the median of posterior p(s|x). Ex. 3: loss function is 0-1 error: 0 if correct, 1 all other choices Optimal policy is to report the mode of posterior p(s|x).

34. An Example: Binary Hypothesis Testing Decision Problem: xx … x1x1 s x2x2 x3x3 Incorporate evidence iteratively (Bayes’ Rule):

35. An Intractable Solution in Practice Stopping Cost Too hard!

36. Generalize Loss Function We assumed loss is linear in error and delay: Wald also proved a dual statement: Amongst all decision policies satisfying the criterion SPRT (with some thresholds) minimizes the expected sample size. This implies that the SPRT is optimal for all loss functions that increase with inaccuracy and (linearly in) delay (proof by contradiction). (Bogacz et al, 2006) What if it’s non-linear, e.g. maximize reward rate:

37. Examples Ex. 2: loss function is absolute error: Optimal policy is to report the median of posterior p(s|x). Ex. 3: loss function is 0-1 error: 0 if correct, 1 all other choices Optimal policy is to report the mode of posterior p(s|x).

38. Some Intuitions for the Proof (Frazier & Yu, 2007) Continuation Region Shrinking!