Comparing Training Methods

1. Comparing Training Methods Evidence for and against the advantages of clicker training

2. Clicker training is misunderstood and misused • Assume that just use the clicker • Not realize that have to condition the click: is a conditioned reinforcer, not a primary reinforcer • Have to add a cue for the response • Clicker marks the behavior, not elicits it! • Fade the clicker as behavior becomes fluent • No longer need the marker • In a chain, a cue becomes a conditioned reinforcer

3. Cueing Basics: • A fluent cue response is: • Precise: animal performs the behavior exactly as you hadenvisioned • Performed with low latency • Performed with optimal speed • Shows resistant to distraction • Performed at any distance from the handler • Performed for the duration required by the handler

4. Example of a chain: • Each cue serves as: • A reinforcer for the prior behavior • A cue for the next behavior • A cue to keep going

5. Note that with a behavior chain • Several responses are emitted before a C/T is given • HOW the behaviors are taught is what is different • Shape each step • Often backward chain: start from end and work backward • Work towards • Accuracy • Fluency

6. Very few experimental reports on the value animal training • Just assume it works • Lots of incidental evidence • Does clicker training work? Is it better? • Let’s look at 2 studies

7. Thorn, et al. 2006Training shelter dogs • Shelters are high stress environments • Very little training • Lots of inappropriate behavior inadvertently reinforced • Little time for staff to train • Need for quick but effective training programs • Teach basic manners • Sit is the most basic

8. Three experiments • Experiment 1: • 3 researchers: Handler, timer, “stranger” • Approach dog: record time to sit/start sit • Used verbal “good dog” and a treat • Experiment 2: • Compared verbal vs. clicker • Verbal: waited, said “good dog” and treat • Clicker: waited, clicked, treat • 2 days of training for each: look at speed of training and retention • Experiment 3: • Compared Contingent vs. noncontingent reward • 3 days of 15 minute training with food /verbal reward for sit • 4 post training conditions: • Same trainer, same room • Different trainer, same room • Same trainer, different room • Different trainer, different room

9. Results • Latency to sit significantly decreased. • All dogs were able to sit within 60 sec

10. Exp. 1, con’t • Mean latency of second session less than ½ of first session.

11. Experiment 2 • Day 1 of training: No difference between clicker and verbal training • Significant decrease in latency to sit • Day 2: verbal training appeared better than clicker training • dogs showed better retention • Dogs showed lower latency

12. Experiment 2

13. Experiment 2

14. Experiment 3 • No differences between groups on first day • Dogs in contingent trials sat significantly longer, and this increased across trials • Was a day x treatment interaction • Dogs in noncontingent condition sat less • Dogs in contingent condition sat more

15. Experiment 3

16. Experiment 3 • No differences in generalization tests • No difference across four test conditions for the dogs in the contingent reward condition • Shows that the dogs generalized the “sit” across new settings and new trainers

17. Experiment 3

18. Summary: • Data show that minimal training in shelter (10-15 minutes over 2 days) works! • Even with novice trainers • Dogs quickly learned sit • Able to retain new response • Why verbal work better than clickers? • Limited supply of verbal reward in shelter, so more valuable? • Negative association or no conditioning to clicker • Incomplete conditioning of dogs to clicker • Other factors?

19. Summary: • Shelter implemented new training policy: • All staff required to have dog sit when moving dog, feeding dog, interacting with the dog • Dog exposed to continuous training across settings • Found that other behaviors also affected: • Decreased inappropriate responses such as barking, stress responses, jumping on cage • Better response to the dogs; increased adoption probability

20. Ferguson and Rosalez-Ruiz (2001) • Trailer loading = critical horse behavior • Horses often not like trailers • Small, dark, confined • Aversive methods often (usually) used • Too much negative reinforcement and punishment, which often escalates to increasingly aversive treatment

21. Horse Loading Behaviors • Back up • Move Forward • Turn left/right • Step up • Loading problems = leading problems • Horse not going where led • Balks, turns away, etc.

22. Method • 5 horses with poor loading history • 2-horse straight load step-up trailer • Butt chain instead of butt bar • Side windows and rear doors left open • White inside and outside • Railroad tie used as extension of trailer deck • Target: Red pot holder • Reinforcers = typical horse treats • 15 min training sessions

23. Method • Baseline compared to training • Loading behavior chain: horse approach trailer: • Within 3 meters (about 10 feet) • Within 1.5 meters • With head/neck in trailer • With front legs in trailer • With ½ body in trailer • With 3 legs in trailer • With 4 legs in trailer; less than 5 sec • 4 legs in, allowed butt chain to be fastened, door to be closed.

24. Behaviors Recorded: • Inappropriate responses/stress responses: • Amount of horse in trailer (using 8 step chain) • Freezing • Head toss • Standing • Turning • Loading: • Getting into trailer (less than 5 sec) • Loading and staying in trailer • Number of prompts • New leads (re-approaches) • Latency to respond to cue: • Within 5 sec • Greater than 5 sec • No response • Also obtained interobserver agreement (IOA)

25. Procedure: • Baseline: 1 day of repeated 5 min baselines • Target training: 2 days; 20-30 trials/day • Touch target • Criteria: 80% of prompts • Trailer training: • Trials to touch (just inside trailer) • Upon entry, lead back to start and another trial • Started at each horse’s baseline distance • Added then faded trailer extension • Trained to load on left/right sides • Added limited hold with Fancy: gave several steps to move forward • Multiple baseline design across horses

26. Results • All horses learned to target during first training session • First session: average of about 60% accuracy • Second session: average of 80% • Sammy took 3 sessions, but reached 90%

27. Red’s data: • Red shaped quickly: • reached criteria at each step of the chain before moving on to next • Change in setting disrupted behavior, but recovered criteria quickly

28. All horses able to be loaded • Baseline: no horse able to get beyond step 4 • After initial target training: • Red: step 6, some 7 • Penny/Shadow: steps 5,6 mastered; some 7,8 • Sammy: steps 3 to 5 mastered; some 6-8 • Fancy: 5 to 6 mastered • When added extension: all but Fancy reached criterion performance • Fancy outwitted researchers: could stretch to touch target even with extension • Had to add the limited hold condition to outwit him.

29. Combined horse data • All horses maintained loading behaviors when extension removed • Loading left and right and new trailers produced some disruption but quickly recovered • All reached 90%

30. Inappropriate Responses • Most common: • Standing • Turning • Head toss • Immediately decreased with training • Note: not targeted • Suggests these are stress responses

31. Leads and Prompts • During baseline: Few leads and LOTS of prompts • During training: • Fewer prompts • Leads were about 1:1 with prompts

32. Summary • Why is clicker training better? • Faster responding with fewer disruptive responses • Fewer avoidance responses • Happier horses and trainers • Changes in loading procedures (e.g., left vs. right position, new trailer) easily dealt with and overcome

33. Summary • Target training decreased inappropriate responses secondary to increasing trailer loading • Target training established stimulus control • This allowed stimulus control during situations which usually elicited problem behaviors • Horses so busy focusing on target that they ignored poisoned cues

34. Conclusions • Did the Thorn study show that clicker training was not as good, or that in limited use with naïve users it was less effective? • Did Thorn, et al., measure inappropriate responses or the effects of the two training methods? • Which do YOU think is more effective, and why?