Dog Play and Prediction Error

Sep 6, 2019 | Animal Behavior, Learning Theory, Training

Dog play is comprised of “ritualized aggression” © Susan Nilson

Dog play is comprised of “ritualized aggression”, and aggression is rooted in fear. In ritualized aggression, there are varying degrees of fear/stress/aversions, along the spectrum from slight stress to possibly full blown fear, that results in a fight or bite or an attempt to flee.

The fun, enjoyable parts of the “dog play stress spectrum” are positively reinforced, and intrinsically so, by their very nature. These are behaviors that are reciprocated and or result in play being accepted by the other dogs. Play bows, shoulder rolls, invitations to be chased, barking, reciprocal mouthing, and mutual give and take pining and physical exchanges are all signs dogs are enjoying the play.

The aversions, teeth, body checks, too much being pinned and being chased too much, are on the negative reinforcement side of the equation. We see that play behaviors straddle both sides of the spectrum, as dogs are typically working in micro seconds to avoid these slight stresses, by way of movements issued to avoid aversions.  

Some play behaviors are parallel positive and negative reinforcements, toggling at various speeds. The duration of the play will factor greatly in how this R+/R- toggle plays out. A mouthing behavior that becomes too hard, or a pinned dog that is having trouble with breathing could in a nanosecond result in “too much stress” and thus the play can “tip over” into fearful aggression, as dog play may be on the edge of stress and play becomes too much for a dog. This is why it is crucial for humans to understand how to referee dog play based on meta signals from dog body language, having an accurate history of the dogs in play, and proactively shaping the dog play events to reduce micro stresses along the way with training, counte conditioning and breaks in play by way of shaping dog play and refereeing.

Dog play, like human sports, is comprised of physicalities that may result in stress, fear or pain. This is why “ritualized aggression” has a component of “potential” fear, and this is where prediction error can be manipulated by humans.

Dogs are giving and receiving bites, tackles, bumping, pinning, chasing, barking and vocalizing along the whole spectrum of whining, barking and growling. All of these components can lead to fearful determinations during play and play can “tip over” into bonafide aggression. This is especially true with puppies, and dogs that have little to no play history with one another.

When humans are refereeing dog play with deft and gentle tactile instructions, and verbal disengagement training cues, mitigating the toggle between positive reinforcements – (play being well received and reciprocated) and negative reinforcements – (dogs trying to avoid stress/fear/pain) associated with play, bite/chase/tackle/pin, or fleeing by getting away),the dogs will have less stress, hence less chances to “get it wrong”, changing the predictive value  via proactive refereeing of dog play, and thus less chances of dogs fighting due to play becoming too rough. Humans need to be 1000% engaged in the dog play so they can mitigate the stress and teach the dogs how to play and that they are safe.

There are many micro associations being made by dogs at the “snout level” during vigorous play, and as dogs are toggling between the reinforcers of play being an intrinsic positive reinforcer, good reciprocal play (R+) and avoiding play (R-) running away, fleeing, avoiding bites, or tackles, or attempting to return the bite/chase/pin/tackle sequences, in order to “defend”, dogs need help by way of momentary separations and or short breaks in play for stress to stay at acceptable levels. Some dog trainers call these separations “splits”.

Dopamine plays a large role in the circumstances that involve learning, mainly prediction error, and salience of stimulus. Dogs find other dogs very salient, and during play, especially with puppies and dogs new to each other, dogs are learning at rapid rates about both what is reinforcing and what is not reinforcing. Dopamine modulates the expression of learning in performance, primarily response vigor and memory retrieval. Dopamine is considered key factor in the reward related processes that may drive the learning of “extinction contingencies”. This relates to dog play in the sense that when stress is mitigated by way of human refereeing, and dogs are learning what is safe, by way of orchestrated disengagements, this disrupts stress responses, and dogs are less stressed by what is associated as not potentially safe, dogs learn quicker and thus; extinctions, or reductions of unwanted behaviors, and fears or stresses related to dog play are more easily achieved both naturally via play and conditioned via human interventions.

Prediction Error is the processing of stimuli and events and the differences between received/perceived and predicted rewards. This is crucial for evolutionary benefits of social interactions.

Rewards are a big part of learning. Dogs are always assessing “predictive value’, this is something widely known, and prediction error is involved in both associative and operant learning.

It works like this, there is a distinguishing between a prediction about a future reward, or no prediction, which is just a poorly defined one. Then there will be a comparison of the reward and the prediction. When the reward is either better than or equal to or worse than what was predicted, future behavior will change based on the differences between the reinforcements and the predictions.

  • If the reward is different from the prediction, a prediction error exists, and the brain updates the prediction and behavior changes accordingly.

EX – Dog predicts a play bow from another dog, but instead is tackled and pinned. Next sequence the dog sets up for a tackle and pin by adjusting their play sequence in some way towards the other dog.

  • If the reward is better than, positive prediction error,which all creatures want, the prediction becomes better and behavior to reach those rewards/reinforcement will increase.

EX Dog predicts a body check from a larger dog, but instead the larger dog self-handicaps and rolls onto their shoulder. Dog reciprocates with self-handicapping.

  • If the reward is worse than the prediction, negative prediction error, which no creature wants, the prediction becomes worse and behavior that maps to avoidance will increase.

EX Dog receives a hard bite in play, not a damaging bite, but it hurts, next engagement with the other dog will adjust behavior to avoid being mouthed/bitten.

Whether the prediction is better or worse, or new and unexpected, learning is taking place.

Much of learning is accomplished by making mistakes. Although mistakes are something all creatures are striving to avoid, they will occur, and actually help in learning about how to obtain more rewarding/reinforcing behavior. The goal of dog play shaping and refereeing is to minimize the mistakes and the stress by way of proactive training with disengagements.

If there are no further mistakes, the behavior will not change until the next error. Proper human refereeing of dog play with allow for quicker recovery and bounce back from stressful events in dog play. This allows for a more even flow of dopamine to rise and fall without dramatic increases or decreases. Thus, the predictions about play will be more accurate, and simultaneously new behaviors are shaped.

This applies to both learning about obtaining rewards/reinforcements and movements/sequences (operant) to avoid stress/negative outcomes (associative).

Neurological aspects of dog play There are 5 neurochemicals that are interacting when bodies are in motion, and flowing in sequences, be they play, work, or problem solving. Dog play encompasses all of these neurochemicals.

Serotonin, Endorphins, Norepinephrine, Dopamine and Anandamide.
Anandamide is known as the “joy” chemical and helps with lateral thinking,or bridging the gap between disparate ideas.

These chemicals “jack up” the ability for pattern recognition, prediction error, and linking familiar patterns. These chemicals lower the “signal to noise” ratio and help detect more patterns. (better prediction error).

Neurochemicals “tag” experiences and save predictors, teaches the animal about the event and the experiences.

When all of these chemicals are flowing and the dog is feeling safe, plays becomes autotelic, highly motivating, and end to itself.

Dopamine and Rewards

Dopamine neurons are located just behind the mouth in the midbrain. Dopamine axons extend into the amygdala, frontal cortex, striatum, and several other brain regions.

The action potentials of dopamine neurons induce learning and approach behavior. This is a big part of puppy play in the early stages of learning as seen in the “approach avoid” sequences where puppies approach another dog or puppy and approach to solicit play then retreat and “avoid” being chased, but actually in many cases instigate the chase.

The higher the reward/reinforcements, the stronger the dopamine response. Dopamine not only increases when an animal receives a reward, but also when there is stimulus that predicts a reward.These reward predicting stimuli are conditioned responses. Dopamine neurons treat predictive rewards in a similar way to tangible rewards.

Predictive stimuli allows animals to “plan ahead” to make informed choices. Dopamine neurons  provide information about past and future rewards. This is helpful for learning and decision making.

Wolffram Schultz, MD, FRS states, “Dopamine neurons are not reserved for any old reward, dopamine neurons respond to only to rewards that differ from their prediction. The dopamine response is then transferred to the next predicting rewarding stimulus and ultimately back to the first predictive stimulus”.

The longer the time between the first stimulus (sight/scent/sounds of dogs) and the final reward (contact with dogs), the smaller the dopamine response, as subjective reward values become lower with delays. This is a phenomenon known as temporal discounting”. He goes on to state…

The response to reward-predicting stimulus itself depends on the prediction of that stimulus, in the same way as the response to the reward.

Thus dopamine neurons respond to reward-predicting stimuli in the same way as to rewards with only slightly less intensity, which allows animals to use predictive information for teaching even earlier about stimuli and actions.

Schultz goes on to say “dopamine signals are useful for learning long chains of events”. This helps with processing of prediction errors, as not all of the information about the event needs attention, this allows for more efficiently processed information, and in the case of rewards, dopamine neurons are excited with larger than predicted rewards/outcomes.

Dog play encompasses many “long chains of events” and there are many sequences that may not be predicted accurately among dog just learning to play – puppies or dogs that are new to one another, hence when humans proactively shape the sequences in dog play, the dogs develop more accurate predictive value. It’s not that humans prevent play, what is being doe when dog play is shaped and refereed properly is, dogs are learning what works for reinforcements and that they are safe from stress levels becoming increased, which causes a drop in dopamine and a rise in glucocorticoids. In addition, serotonin helps with control over aversive events, and builds resiliency with stress.

Dog Play, what is occurring? Dog play is as individual an event as the dogs and the play styles either formed or forming in the case of puppies, it is a fluid give and take, and what humans want is reciprocaland has equal give and take. The analog would be human contact sports like Basketball or Rugby where there are lots of running sequences and physical interactions, the participants are not fearful, but cautious and implementing avoidance behaviors. The goal in dog play is less stress, no fights or bites.

Size and weight differentials are the biggest challenge. They can be worked with as long as the humans are aware of how to mitigate the size and weight differentials. This is the number one cause of issues in dog play along with too much duration which leads to stress.

There is no formula or training plan for size and weight differentials, as the dogs ability to read the size differences and the play styles that the dog has currently will be the deciding factor in how to approach the size and weight differentials in play. Some weight and size disparities may be too great to consider off leash play. It is a moment to moment event and a case by case assessment of the dogs involved.

Mouthing, nipping, chasing are at some points part of most dog play. These are all normal and depending on the dog’s history and age, should be a parallel concern along with size and weight differentials when refereeing dog play. If the mouthing, nipping and chasing is reciprocal, it can still cause some stress. It is often not the first few teeth that make contact but dozens of “mouthing events” taken place over X amount of time that cause a build-up of stress and a prediction error that could have been corrected had the humans involved in the play been aware of how to mitigate durational aspects of play and reduce stress.

“Ground Game” play this is often overlooked as the dogs are “paying nice”, and while that may be the case as laying down is usually a the byproduct of one or both dogs self-handicapping and the dogs reading it, however there is quite a bit of mouthing that goes into ground game play, and pinning can occur, which may lead to dogs struggling for air, and there may be exposed bellies and errant paws that can find their way onto a dogs stomach and case stress. This is easily mitigated by way of humans having a close watch over the dogs and implementing micro or macro adjustments where needed to reduce stress.

The Positive Reinforcement Negative Reinforcement Toggle Of Ritualized Aggression.

The antecedent of negative reinforcement sequences is an aversive, the behavior that follows is something to avoid /combat that “aversive”, the dog is looking to avoid, or decrease the aversion that is occurring. In the case of dog play, it’s built on aversive avoidances and reinforcing physicality that toggle back and forth.

There are micro aversions involving teeth, chasing, pinning, body checking, and vocalizing. All of this plays into the dogs “reading” each other’s play signals.
Stress can build up during play when humans do not proactively referee the play and the stress can cause dog play to tip over into a fight or dogs will shit down and “tap out” and those deference signals that show humans a dog is done with play helps to reduce stress and avoid negative associations.

It is often not the first tooth that lands too hard on the other dog, but the 20thtime, and thus the negative reinforcement was not achieved by way of some avoidance movement, and now the dogs have stress past an acceptable point.

Helping reduce stress during dog play by way of tactile and verbal disengagements creates negative reinforcement sequences by way of disengagements before too much stress builds up.Humans are thinking for the dogs, and implementing breaks in play before it is needed.

This is crucial for puppies and dogs that are just learning about their play styles with new dog friends. Dogs should not be allowed to fully “figure it out”, they need guidance and structure so they have the most rewarding experiences during play. Dogs are making prediction errors on both sides of the equations for both potential reinforcing and potential avoidances.

Dogs move 7 – 10 times faster than humans, however humans think faster and with more clarity and accuracy than dogs do, especially when there is a hectic environment to contend with. Humans process data exponentially faster than dogs. When humans implement the negative reinforcements – avoidance sequences earlier, before there is an inordinate amount of stress, the humans shape the sequences for better prediction error and prediction value, thus the dogs learn faster they are safe.

Prediction error is crucial in learning, and what is most prescient is the dogs having predictions be better than expected.

Predication Error, Dopamine and Dog Play

  • If the reward is different from the prediction, a prediction error exists, and the brain updates the prediction and behavior changes accordingly.
  • Dog play is occurring at lighting speeds many times, and there may be more that 2 dogs in the play group, this means that “predictions” are going to be somewhat fuzzy, especially for puppies and dogs that are getting to know one another. By refereeing the dog play and proactively changing predictions about potentially stressful events both macro (dogs charging at a dog to solicit play and humans block the oncoming dogs to avoid the dog being bombarded is changing the prediction) and micro events (dogs laying together playing and mouthing humans mitigate teeth and paws for less potential stress), the dog will start to create new predictions based on the humans mitigating the dog play.
  • If the reward is better than, positive prediction error,which all creatures want, the prediction becomes better and behavior to reach those rewards/reinforcement will increase. This is the byproduct of dog play refereeing as described above, often times in puppy classes we see puppies that were fearful at the start of play, then through counter conditioning, proactively refereeing to decrease stress and or using management with gates and implementing counter conditioning, the dogs update their predictions and decide to enter the play group fully. Or we pair some of the more relaxed dogs with each other, we change the environment, and thus the predictions change, and the puppies that were fearful of play with the energetic dogs are now playing with the puppies that are less energetic, the dogs “update the predictions”.
  • If the reward is worse than the prediction, negative prediction error, which no creature wants, the prediction becomes worse and behavior that maps to avoidance will increase.
  • This is how play tips over into a fight or dogs shut down and hide from dog play, or perhaps hard bites are landed. Puppies will have “fear periods” and in general will have the least amount of play history, and many times puppies are intrinsically apprehensive towards other puppies, as they do not move like the older dogs they have played with, or perhaps they have not had play outside the litter mates. It is crucial for humans to proactively referee and shape the dog play so dogs do not have worse predictions, but better than expected.

The human variable in dog play : How dog fights are avoided and dogs learn how to play better.

In over 10,000 play groups since 2006, comprised of all manner of dogs, puppies, rescue dogs, my own dogs, dogs that boarded with me, there has never been a dog fight or seriously bitten dog. Not one dog fight in over 10,000 dog play groups, not one. All “tip overs” were at an acceptable and expected level of aggression. The vast majority of these play groups are on film, from 2012 – present all dog play groups I conducted are on film.

Dog play is comprised of ritualized aggression much like human sports, there will be rough and tumble events in dog play, and while there may not always be a big mouthing, chasing, pinning exchange, there could be, even if for the first few minutes, or sudden bursts of play can occur, and whether dog play is chaotic or calm, there can still be stressors that build up and lead to dogs feeling unsafe.

Humans by staying 1000% engaged in the dog play and having an accurate history of the dogs in the play group, and employing a force free non aversive referee style to shape the play, this is how to avoid dog fights and or extreme dust ups in dog play. Dog play shaping and refereeing are also how dogs are guided to more acceptable play styles when size and weight differentials are a concern.

Dog Play Refereeing and Shaping takes humans to be mentally and physically fit. The reason dog fights break out during play is people are not taking accurate histories and not staying engaged in the dog play to reduce stress.

There is a very misguided notion that during dog play “dogs can figure it out”, that is what dog fighters do, they let dogs figure it out. Humans need to learn about dog play signals, dog stress signals in play, and develop a referee style that integrates with their physical abilities, and really know the dogs they are working with in a dog play group or duo by taking accurate assessments of the dogs play histories and watching video. Dogs can figure out some of the proceedings, but humans need to take a practice role and shape the play for best results.

Dog play occurs on the Nano-Second and there are lots of events at the “snout level”, millions of bits of behavioral data being processed, many times these behavioral data bits are related to cellular DNA and survival triggers, dogs may get it wrong or not have sufficient historical info on the dogs they are playing with to determine what is the best course of action, and a dog fight can occur.

When humans are focused, humane, and proactively seeking to reduce stress, the chances of a dog fight breaking out during play are incredibly low, as in the 1% category. After all that is the macro goal of all dog play groups or duos, no fights or concerning bites.

By taking a proactive role in dog play and refereeing, humans are creating better prediction error and that maps to safer dog play with less stress.

References

Sapolsky, R. (2011). Human Sexual Behavior III & Aggression I. Available at: https://www.youtube.com/watch?v=JPYmarGO5jM

Sapolsky, R. (2011). Aggression 2. Available at: https://www.youtube.com/watch?v=wLE71i4JJiM

Sapolsky, R. (2011). Aggression 3. Available at: https://www.youtube.com/watch?v=EtVfoIkVSu8

Schultz, W. (2016). Dopamine Reward Prediction Error Coding. Available at: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4826767/

TEDx Talks. (2016). How to open up the next level of human performance | Steven Kotler | TEDxABQ. Available at: https://www.youtube.com/watch?v=7xnbUT3rOvQ&t=165s

Urban Dawgs. (2011). Shaping Dog Play. Available at: https://www.youtube.com/watch?v=BFyFoz-Q6wA&t=96s

Urban Dawgs. (2019). Prediction Error and Dog Play. Available at: https://youtu.be/3MtPVmsDWR0

Urban Dawgs. (2019). Dog Play Shaping & Refereeing Puppy Kindergarten. Available at: https://www.youtube.com/watch?v=Z7FTUdqb_PU

Urban Dawgs. (2019). Dog Play Narration Refereeing and Play Shaping. Available at: https://www.youtube.com/watch?v=IdFwMSmrN5w&t=12s

Urban Dawgs. (2019). Setting Up Dog Play Body Language and Meta Signals. Available at: https://www.youtube.com/watch?v=aud3eI6urqc&t=9s