Sun Apr 16 20:13:21 PDT 2006

Virtual Skinner Box

I was surfing Karen Pryor's web site and ran across a wonderful training resource called Practice Your Clicker Timing. It's a Macromedia Shockwave application that lets you practice your clicker training skills online by simulating a pigeon in a Skinner box.

The page is densely packed with information, and what may get lost in the information overload is that you need to actually click on the virtual pellets to dispense a treat; clicking elsewhere has no effect at all. It took me a few minutes to figure that out, even though this is mentioned in the first paragraph of explanatory text. (Doh!)

Anyway, while I don't know how accurate the simulation really is, I played with it long enough to realize that it does help the casual trainer to improve their efficiency and to understand some of the trade-offs involved in various reinforcement schedules. After some early failures, I managed to eventually get the pigeon to peck the target at rates as high as 588 repetitions in a 30-second period.

Interestingly enough, once I reached rates that high, using less reinforcement reduced the overall rate of performance, but increased the total effort (or "pecks per treat") that the virtual pigeon was willing to put in. This would appear to be a good example of how to use extinction bursts and variable reinforcement schedules to increase a desired behavior.

For example, I doled out 22 treats over 30 seconds at the pigeon's peak rate, for an overall reinforcement schedule of 26.7 pecks per treat. When I cut the reinforcement rate in half, to 11 treats over the next 30-second period, the pigeon only provided 491 pecks, but at an astonishing 44.6 pecks per treat!

Looked at another way, I improved the potency of the rewards I was handing out by a whopping 167% at the expense of total work effort provided by the pigeon, which dropped by about 16.5%. Whether it is intrinsically more valuable to maximize efficiency of rewards or total work effort in real-life applications is, of course, a complicated question to which I have no immediate answer.

It's also worth noting that the simulation doesn't take external cues into account. What it's simulating is the effects of the trainer's timing and use of reinforcement rates on an animal's spontaneous behavior. This does serve the purpose of reminding the trainer (especially those from a command-based background) that behaviors can be taught in a reliable, solid way long before a cue is added. However, recognizing that the simulation is about shaping, rather than cueing, is critical to getting the most out of the experience.

Don't let all this pseudo-scholarly analysis scare you away from the site, though. It's worth repeating that this is a great tool for beginning clicker trainers. The fact that one can use it to explore theory and advanced techniques is a plus, but certainly not a prerequisite.

I highly recommend it to anyone interested in clicker training. Try it out, and let me know what you think.

Posted by Todd A. Jacobs | Permalink