1. Quantizing Behavioral Heterogeneity Jon Beckham 11/21/02

2. Papers to Cover  “Measuring Robot Group Diversity”, Balch  “Design & Evaluation of Robust Behavior- Based Controllers”, Goldberg & Mataric  “Symmetry in Markov Decision Processes and its Implications for Single Agent and Multiagent Learning”, Zinkevich & Balch

3. Quantizing  “Measuring Robot Group Diversity”, Tucker Balch

4. Purpose  To suggest a standard way of quantitatively measuring diversity. Allows for more accurate, effective analysis. By establishing a standard metric, we can establish a baseline for comparison.

5. Sources  Simple Social Entropy Adapted from Shannon’s Information Entropy  Behavioral Difference Quantitative measure between different robots.  Hierarchic Social Entropy Combination of the above.

6. Diversity  To quote Tucker, who quotes Webster… di verse adj 1: differing from one another: unlike. 2: composed of distinct or unlike elements or qualities.

7. The Discrete Approach  Assume robots are either alike or different; thus assume subsets of identical robots.

8. Simple Social Entropy  First, some notation: R is a society of N agents, thus R = {r 1, r 2 …r N } C is a classification of R into M subsets c i is an individual subset of C Thus C = {c 1,c 2 …c M } p i is the proportion of agents in the ith subset. Thus, the sum of all p i is 1.

9. Social Entropy’s Requirements  Continuous (H must be continuous in p i )  Monotonic (H must be monotonically increasing function of M)  Recursive (H must be weighted sum of H of subsets)  H = 0 when system is homogeneous  H is maximized when all p i are equal for given M  Any change to p i to approach greater equality increases H.

10. Thus…  H(X) = -K∑ M i=1 p i log 2 (p i )  REMEMBER THIS!  Also know that it’s the only equation to satisfy the first three properties (as proven by Shannon in his information entropy work).

11. Limitations of Simple Social Entropy  Loses data by munging p i and M into single value.  Only works for discrete systems.

12. What About C?  The classification into subsets… Taxonomy Clustering

13. More on Taxonomy  Classification at varying levels through a “dendrogram”.

14. Which Brings Us To Hierarchic Social Entropy  Simple Social Entropy is only a “snapshot” at a particular level of clustering.  To achieve a continuous metric, we use a plot of entropy at all taxonomic levels.  Good because it gives data at all clustering resolutions, putting to rest the clustering issue.

15. Another Formula  This time for hierarchic social entropy.  S(R) = ∫ 0 ∞ H(R,h)dh

16. Branching the Taxonomy?  How to get that pretty 2D mapping… Evaluation Chamber? In real world, this requires:  Fixed policies  Mechanically Homogeneous  Policy is reflected directly in overt behavior

17. Placing Numerical Value on Behavioral Differences  More notation i is a robot’s perceptual state a is the action (behavioral assemblage) selected by a robot’s control system based on the input i. π j is r j ‘s policy; a = π j (i) p i j is the number of times r j has encountered perceptual state I divided by the total number of times all states have been encountered

18. Simple Behavioral Difference Metric  Continuous D’(r a,r b ) = 1/n ∫ | π a (i) - π b (i) | di  Discrete D’(r a,r b ) = 1/n Σ i | π a (i) - π b (i) | (1/n is normalization factor)

19. Behavioral Difference  Continuous D’(r a,r b ) = ∫ (p i a + p i b )/2 | π a (i) - π b (i) | di  Discrete D’(r a,r b ) = Σ i (p i a + p i b )/2 | π a (i) - π b (i) |

20. Definitions  Absolutely behaviorally equivalent Iff two robots select the same behavior in every perceptual state.  ε-equivalent if D(r a,r b ) < ε.  ≡ ε indicates ε-equivalence  A group of robots, R, is ε-homogeneous if for all r a,r b in R, r a ≡ ε r b.

21. Experiments (briefly)  Multiforaging Behaviors wander stay_near_home acquire_red acquire_blue deliver_red deliver_blue Perceptual Features red_visible blue_visible red_visible_outside_homezone blue_visible_outside_homezone red_in_gripper blue_in_gripper close_to_homezone close_to_red_bin close_to_blue_bin

22. Methods  Local performance-based reinforcement  Global performance-based reinforcement  Local shaped reinforcement

23. Results

24. Summary  Diversity is good in soccer, bad in simple foraging.  Diversity Globally Rewarded, most diverse Locally Rewarded Shaped, least diverse

25. Conclusions  Diversity as an independent variable  Simple social entropy  Hierarchic social entropy

26. Problems?  Only deterministic policies  Analysis limited to behavioral diversity

27. Applying  “Design and Evaluation of Robust Behavior- Based Controllers”, Dani Goldberg and Maja J. Mataric

28. The Goal  To design multirobot controllers that: Exhibit group-level robustness to robot failures and noise. Are easily modified.

29. Focus  Simple Foraging

30. Controllers  One Homogeneous  Two Heterogeneous Pack Caste

31. Homogeneous Controller  Act concurrently and independently.  Behaviors Avoiding Wandering Puck Detecting Puck Grabbing Homing Boundary Buffer Creeping Home Detector Exiting Reverse Homing Heading

32. Heterogeneous Pack Controller  Uses temporal arbitration SPST → SPDT  Dominance hierarchy based on capabilities or arbitrary assignment  Only one robot can deliver a puck at a time  Same controller as homogeneous, but uses ‘message passing’ to figure out which robot should deliver first.  Uses communication to determine failed or active.

33. Heterogeneous Caste Controller  Uses spatial arbitration SPST → DPST  Robots are differentiated into sub-groups or castes  Act concurrently and independently, but in different regions of the task space  May have heterogeneous behavior in addition to spatial heterogeneity  No reliance on communication (Not implemented, but communication could be use to balance caste ratios in case of failure.)

34. Interference Graphs  Homogeneous  Heterogeneous Pack  Heterogeneous Caste

35. Analysis Metrics  Inter-robot collisions  Distance traveled by each robot  Time-to-completion

36. Statistics… Goldberg & Mataric: “We have performed hypothesis tests using Student’s t, 1-factor analysis of variance (ANOVA), and 2-factor ANOVA, in order to verify that the differences between the results of the implementations were in fact statistically significant.” Tucker:

37. Results

38. Conclusions  Attempted to apply Balch’s SSE and HSE, but because of vague definitions no clear conclusion could be reached.  Attempted several calculations, but no conclusive relation to performance.  Partly because no best controller.

39. Flaws  Use of communication in Pack controller, but nowhere else. Allowed pack controller to keep track of state of other robots (working or non-working).