by Brian Tomasik
First written: 25 March 2013; last update: 17 Mar. 2017
It's easier for an agent to learn with a combination of negative and positive reinforcement than just one or the other, and this probably explains why organisms feel both suffering and happiness. Because the painfulness of dying is not necessarily subject to constraints on learning fitness tradeoffs, it can vary widely in how bad it feels, from being quick and easy to, more commonly, indescribably awful.
Note: This article was written in pieces over time and is therefore not very cohesive.
See also Shlegeris (2016).
- 1 Summary
- 2 Question
- 3 Seeking vs. avoiding
- 4 Learning
- 5 Pleasure as approval?
- 6 Maximizing vs. satisficing
- 7 Reward and punishment as separate brain systems
- 8 Painfulness of dying
- 9 Are happiness and suffering merely relative?
- 10 Acknowledgements
Why aren't animals motivated only by gradients of bliss or gradients of agony? Why do emotions sometimes have positive valence and sometimes have negative valence?
Seeking vs. avoiding
Some things animals are supposed to avoid (fire, predators, bodily injury, etc.), and some things they're supposed to seek out (food, shelter, mates, etc.). Avoiding something is easier for a body to do by making that specific thing painful than by making everything else in the universe except that specific thing more pleasurable. Yew-Kwang Ng claims that emotions are metabolically costly, so the body aims to conserve on their use. A parallel argument goes for pleasure: It's easier to make the specific thing to seek pleasurable than to make everything but that thing painful.
That said, there are subtleties. When you get hungry, you don't just have a lack of pleasure, but you have an active discomfort. This is sort of like "make everything painful except eating." The same can be true for other pleasures, including addictions: Sometimes it's not that the drug makes you feel good -- it's that doing anything other than taking the drug makes you feel bad. So maybe it's not inconceivable after all to imagine life as gradients of discomfort that are more or less relieved by doing certain things. Probably depressed people do feel this way. And those with a high hedonic setpoint may be closer already to gradients of bliss, although there must be some exceptions for really severe pain.
It turns out there's a large literature on this topic. For example, a number of references are cited in "The effects of positive versus negative feedback on information-integration category learning."
A long history of research has investigated the relative efficacy of positive and negative feedback. For example, early two-choice discrimination-learning studies with rats found that punishment-only training caused faster learning than reward-only training (see, e.g., Hoge & Stocking, 1912; Warden & Aylesworth, 1926). The first human studies, which used simple two-choice rule-based category learning tasks, also found that negative feedback was more effective than positive feedback (see, e.g., Buss & Buss, 1956; Buss, Weiner, & Buss, 1954; Meyer & Offenbach, 1962). More recently, however, Frank, Seeberger, and O'Reilly (2004) reported that dopamine replacement medications reversed this effect in Parkinson's disease patients (i.e., positive feedback became more effective than negative feedback). Several researchers hypothesized that the more commonly observed negative feedback advantage occurs because positive feedback is less informative than negative feedback, at least in two-choice tasks (Buchwald, 1962; Jones, 1961; Meyer & Offenbach, 1962). The idea is that negative feedback informs the participant that his or her hypothesis was incorrect and also signals which response was correct (i.e., the other response), whereas positive feedback signals only that the response was correct (i.e., the hypothesis might have been incorrect, but, by chance, the response was correct). With more than two categories, negative feedback loses some of this advantage. This information asymmetry hypothesis was supported by results from a four-category study that found no difference between negative and positive feedback (Buss & Buss, 1956).
It's interesting to observe how Popperian science is based on negative feedback (falsification).
Some more discussion from the world of animal training. "Dog Training: Positive Reinforcement vs. Alpha Dog Methods"
Sylvia-Stasiewicz admits results can come slower with purely positive reinforcement, but says the method has even saved so-called "death row dogs" who some thought impossible to rehabilitate.
When clicker training was first being applied to horses, a lot of the training was done using only positive reinforcement and the model of dolphin training was the one most people used. This was true of dog training too. A lot of early dog training emphasized free shaping and the animal was at liberty and able to choose to participate in the training or not. This worked very well for training many types of behaviors but was not quite the same as training a horse to be ridden. Unless horse people wanted to throw out everything and start over, this meant horse people had the interesting challenge of figuring out how to combine a positive reinforcement based training system with a negative reinforcement based training system.
(I do not necessarily endorse the practices in these articles.)
Similar considerations about positive-vs.-negative feedback apply for human learning and working.
In his comments on Marvin Minsky's The Emotion Machine, Len Schubert provides a helpful illustration of why it's hard to turn negative reinforcement into positive reinforcement:
If one imagines trying to build a robot that learns strictly through positive reinforcement, one can see the difficulty: for instance, suppose we designed the robot so that if it receives a leg injury, it will get considerable pleasure from treating the leg with great caution and care, until it is healed (or repaired). Wouldn't that lead to appropriate responses to injury? Well no -- it would probably try to get injured, so as to enjoy the feeling of caring for the injury! Can we do better by designing the robot to get pleasure from injury *avoidance* -- i.e., it gets positive reinforcement whenever it perceives that it *might* have been injured, but didn't get injured? Well, it would *still* seek out dangerous situations, since otherwise it'll have no sense of having avoided injury! So perhaps we want to build it so that the safer from injury it feels itself to be, the happier it is. But then, what would prevent it from neglecting an *accidental* injury? This may be solvable, but it doesn't look easy... or can we just "shift the origin" on the scale of negative and positive feelings, so that all feelings are just more or less positive, never negative? Then even an injured creature or robot would be feeling not-too-bad, yet would be striving strenuously to get help and/or take measures to promote healing of the injury, wouldn't it? Or would it?? If the shift in origin causes no behavioral change, then the robot (analogously, a person) would still behave as if suffering, yelling for help, etc., when injured or otherwise in trouble, so it seems that the pain would not have been banished after all!
We should also remember the hedonic treadmill in this discussion. Even things that we seek out can become less positive due to habituation. As mentioned before, sometimes we seek things to relieve the discomfort of not having them.
Pleasure as approval?
One dimension that may distinguish pleasure from pain in humans is that pleasure is a situation that we approve of and want to create more of (both in ourselves and in others); meanwhile, painful emotions are those we disapprove of.
But what about a situation of temporary relief from pain? For instance, suppose you have a headache in which the intensity of pain ebbs and flows but in which pain is always present to some degree. Let's say the intensity of pain ranges from -5 (worst) to -1 (least bad). You might "approve" of a temporary ebbing of the pain (i.e., going from an intensity of -5 to -1) even though the experience at intensity -1 is still net negative.
Maybe we can answer this objection based on a similar point as in Schubert's robot examples. While you might approve of an ebbing of pain within the context of an existing headache, you wouldn't seek to create additional mild headaches with intensity -1 for yourself. In contrast, you would seek to create additional experiences of mild pleasure from, e.g., hearing music.
However, a critic can still complain. Maybe we should actually score the intense headache as -6, the mild headache as -2, no headache as -1, and listening to mildly pleasant music as 0. In that case, the reason you don't seek to create additional mild headaches for yourself isn't that they have negative valence; the reason is just that they're more painful than the "no headache" experience, which has intensity -1. And the reason you listen to music isn't because it has positive valence; it's just less painful than not doing so. Perhaps this sort of view is consistent with Buddhism’s Second Noble Truth: "an agent is suffering whenever it desires or 'craves' to change its state." In other words, any situation where you want to increase your hedonic experience is a situation of suffering. But the numbers I used in this example were arbitrary, and I could have just as easily set severe headache as +1042, mild headache as +1046, no headache as +1047, and music as +1048. Moreover, the Buddhist view doesn't account for the clear heterophenomenological differences in how humans conceptualize positively vs. negatively valenced experiences.
Maximizing vs. satisficing
Tomasik (2014) shows (pp. 18-19) that for a simple reinforcement-learning agent with a fixed lifespan, optimal behavior is invariant with respect to a constant shift of all reward values up or down. This raises the puzzle of what distinguishes pleasure from pain in more complex quasi-reinforcement-learning agents like ourselves. Tomasik (2014) proposed several possible accounts of the distinction between pleasure and pain in humans, and the current piece speculates further on that topic.
In this section, I'll proffer another potential answer to the puzzle. It relates to the Buddhist Second Noble Truth proposal from the previous section -- that suffering can be seen as unsatisfied desires. In this case, both physical pain and lack of craved pleasure count as suffering, while mild contentment, even in the absence of active pleasure, does not. But in the previous section, I gave an example in which we assigned a neutral state of mind (the absence of headache) a reward of -1 and the pleasure of listening to music a reward of 0. According to those numbers, it would seem that the neutral state of mind, with value -1, is "suffering" compared with listening to music, but this doesn't square with our intuition that being in a neutral state can be fine and doesn't count as suffering.
Part of the issue here could be that a simple reinforcement-learning agent is a maximizer: It tries to optimize a (possibly discounted) sum of future rewards, in which case the difference between -1 and 0 matters equally as the difference between 0 and 1. But humans tend to be closer to satisficers: once things are going well enough, we don't always have to seek more and more stimulation, unless we crave it. To a simple reinforcement learner, a neutral state of mind without the pleasure of music is suboptimal and needs to be replaced by the state of listening to music; so to the Buddhist, this agent could still be seen as "suffering" (to some degree) from being in a neutral state when it "craves" something better (i.e., takes actions to get to the more rewarded state). In contrast, the human who's content to have a neutral state of mind for some time is satisfied and is not seeking to maximize pleasure, i.e., is not acting how a simple reinforcement learner would.
In other words, this proposal conceives of pleasure as "a state when things are good enough" and pain as "a state when things aren't good enough, so that you need to take action to fix the problem". In my high-school health class, we read one article that said something like "Pain is a sign that something needs to change", and apparently this idea is a quote.
Being satisfied with one's situation doesn't mean being inactive. For example, you might be motivated to keep eating food even though you're perfectly happy throughout the process. By "being satisfied" I just mean "approving of the way things are going, rather than feeling that a new and different course of action needs to be taken". It's similar to the difference between re-electing an incumbent politician versus forcing her out of office.
A simple reinforcement-learning agent doesn't have this (admittedly fuzzy) distinction between "keep doing what you're doing" versus "change course" (unless it reaches a state of maximal reward, in which case it has no incentive to deviate therefrom). Perhaps it's partly because simple reinforcement learners are pure maximizers that the pleasure/pain distinction doesn't apply as well to them as it does to humans.
Unfortunately, as is the case with my other proposals in this piece, the satisficer account isn't fully, erm ..., satisfactory. While a person can be happy throughout a meal by following the default course of action to "keep eating", we wouldn't say that a person is happy during a mildly painful dental visit by following the default course of action to "keep still until the dental cleaning is done". Possible replies:
- Maybe the distinction is that the dental pain still involves a persistent wish to change one's state, and it's just that that wish has to be suppressed. Similarly, if you're hungry and without access to food, there may be a persistent wish to find food, but that wish is constantly thwarted. In a counterfactual situation where there was a way to turn off the dentist pain, or where food became available to eat, you would take that action, even if in the actual world you're unable to do so. In contrast, someone contentedly eating breakfast doesn't have a persistent, implicit desire for things to change.
- We could also appeal to a more physiological account and suggest that pain usually involves the sympathetic nervous system (fight-or-flight response), while pleasure usually involves the parasympathetic system ("rest-and-digest" or "feed and breed"). However, this distinction doesn't perfectly map onto pain vs. pleasure. For example, one might be very alert while playing an action video game but still be feeling pleasure. It seems we should call the sweaty-palmed video-game player who's on the edge of his seat "satisfied" even though he's very alert and aroused.
Reward and punishment as separate brain systems
One of the most plausible accounts of the difference between pleasure and pain is that the pleasure/pain distinction mainly results from contingent aspects of how our minds are designed. For example, I would conjecture that at a high level, we have some brain systems that represent degrees of pleasure, and different systems that represent degrees of punishment. The difference between pleasure and pain is then a difference in what system is being activated, and our brains presumably have some rough conceptual sketch of which system (pleasure or pain) is being engaged. We then use the labels "pleasure" and "pain" to refer to these different systems.
The following passage from Nakatani et al. (2009) illustrates the idea of different brain pathways for reward vs. punishment (p. 371):
Notably, we also found that different neurotransmitters play critical roles in learning with water reward and saline punishment: octopamine (OA: invertebrate counterpart of noradrenaline) and dopamine (DA) play critical roles in mediating reward and punishment signals, respectively, in olfactory learning and visual pattern learning in crickets (Unoki et al., 2005, 2006). These features are found for kinds of US other than water and saline and in species other than crickets: OAergic and DAergic neurons have been suggested to participate in appetitive olfactory learning with sucrose reward and in aversive olfactory learning with electric shock, respectively, in honey bees (Farooqui, Robinson, Vaessin, & Smith, 2003; Hammer & Menzel, 1998; Vergoz, Roussel, Sandoz, & Giurfa, 2007) and fruit-flies (Honjo & Furukubo-Tokunaga, 2009; Riemensperger, Voller, Stock, Buchner, & Fiala, 2005; Schroll et al., 2006; Schwaerzel et., 2003; but see Kim, Lee, & Han, 2007). Recently, it has been suggested that serotonin plays a role in aversive place learning in Drosophila (Sitaraman et al., 2008); thus, dopamine may not be the only neurotransmitter to mediate aversive signals.
In mammals, midbrain dopaminergic neurons have been shown to mediate reward signals in learning of a variety of stimuli, including visual, auditory and mechanosensory stimuli (Schultz, 2007; Wise, 2004). Transmitters of neurons conveying punishment signals are less understood in mammals, but in some forms of aversive learning, roles of serotonin (Daw, Kakade, & Dayan, 2002), noradrenaline (Harley, 2004) or dopamine (Schultz, 2007) have been suggested. Therefore, learning in mammals and that in insects share the feature of aminergic neurons playing major roles in conveying reward and punishment signals, although the kinds of amines used to mediate reward or punishment signals appear to differ in different phyla.
Of course, this explanation pushes the question back: Why did evolution give us different brain systems for reward vs. punishment? Fortunately, this puzzle may be easier to solve. Maybe, as Yew-Kwang Ng claims, it's inefficient for a brain to always be activating reward or punishment processing. It's probably more effective to only activate them when unusual events happen, like getting injured or having an orgasm. In more ordinary, everyday experiences where nothing significant is happening in either a bad or good way, the reward/punishment systems don't have to fire as strongly, and since these default, ordinary experiences tend to be the most common, the brain on the whole conserves on effort by firing hedonic systems least strongly in the default, ordinary contexts. Maybe this idea of minimizing how often pleasure/pain systems need to fire could be seen as a temporal form of sparse coding: "temporal sparseness ('a relatively small number of time periods are active')". Alternatively, maybe pleasure/pain systems fire all the time, but the default, ordinary experiences activate fewer pleasure/pain neurons than the more intense experiences do. Foldiak and Endres (2008):
If the goal is to maximize sparseness while keeping representational capacity high, a sensible strategy is to assign sparse [lower-effort] codewords to high probability items [such as (I speculate) default, ordinary levels of hedonic experience] and more distributed [higher-effort] codewords to lower probability items [i.e., intense hedonic experiences]. This implements codes with with low energy expenditure, since neural activity consumes energy.
The following figure illustrates this idea. These graphs show the firing rates of hedonic neurons for a hypothetical animal who experiences hedonic neutrality except at time point #3, where it experiences pleasure, and time points #9 and #10, where it experiences pain. The left-side graph shows how this would be coded if hedonic experience has to be represented by the firing rate of a single brain system. Assuming that a faster firing rate always implies a "more positive" hedonic experience, then in order to code for pain using a low firing rate, we need to have a relatively high baseline firing rate so that we have room to go down. But this means the neurons need to fire a lot in ordinary, hedonically neutral situations. In contrast, if we have two separate hedonic systems, one for the intensity of pleasure and one for the intensity of pain, we can fire much less often and thereby save energy. We also now don't have a limit on how intense pain can be (except for the limit imposed by the maximum physiologically possible firing rate of neurons).
The picture painted above contrasts with the situation of a simple artificial reinforcement-learning agent, who receives a reward at every time step (no sparse temporal coding) that's coded by a constant number of bits, such as a 64-bit double (no sparse coding of reward magnitudes). If it was costly to give an artificial reinforcement learner a reward at every time step, we might see programmers converge on designs in which only the unusually good or unusually bad outcomes are reinforced or punished, respectively, with rewards at or close to 0 the rest of the time. In digital agents, rewards and punishments are distinguished by a + or - sign, but biological brains don't have - signs and so need to use a separate brain system to represent negative rewards. Daw et al. (2002), p. 604:
Opponency has a venerable history in psychology and neuroscience. In an implementational form (e.g. Grossberg, 1988), it starts from the simple idea of using two systems to code for events (such as affective events), with one system reporting positive excursions from a baseline (appetitive events), the other system reporting negative excursions (aversive events), and mutual inhibition between the systems and/or opposing effects on common outputs. From a physiological perspective, this neatly circumvents the absence of negative firing rates.
The framework proposed here makes a moral distinction between pleasure and pain look somewhat arbitrary, because pleasure and pain are only contingently represented by different brain systems in biological creatures for incidental neural reasons. Pleasure and pain needn't be separated this way in artificial minds. If true, this (admittedly speculative) point may weaken our attachment to those forms of hedonistic negative utilitarianism that claim a strong difference in importance between all kinds of pain and all kinds of pleasure. Rather, as is consistent with my own moral intuitions, there doesn't seem to be a big moral gulf between minor pains and minor pleasures. Better forms of negative utilitarianism locate the motivation for prioritizing suffering reduction elsewhere, such as in
- the overwhelming badness of intense, unbearable forms of pain, such as during torture (lexical negative utilitarianism)
- the asymmetry between the moral importance of preventing unsatisfied preferences vs. creating new satisfied preferences (antifrustrationism). By focusing on preferences ("arrows" of motivation to change hedonic state) rather than the magnitude of one's current reward level, a (hedonistic-focused) antifrustrationism or Second Noble Truth view avoids the arbitrariness of where the zero point between pleasure and pain is located.
Note that the arbitrariness of the zero point between pleasure and pain is just as much a problem for classical (hedonistic) utilitarianism as for negative hedonistic utilitarianism.
Painfulness of dying
It's interesting to observe that different ways of dying can vary astronomically in their painfulness, even though in evolutionary fitness cost, they're not that different. For example, burning to death feels much worse than freezing to death, and freezing to death feels much worse than dying of a heart attack in the night. Why the discrepancy?
I guess it's because most of our pain sensations aren't targeted toward avoiding death specifically but toward avoiding other unpleasant things. Contact with fire is one of the most deleterious things we can do per unit time, so the pain intensity needs to be really severe. Cold is less damaging, so the negative weight isn't as strong. Evolution doesn't want to discourage hunting in the snow if that's what's needed to bring home mammoth meat.
Presumably one of the main deterrents of death is not pain but fear: Avoiding the predator before it attacks you, avoiding the high cliff, avoiding diseased food or feces, etc. Once the pain starts, it may already be too late in some cases.
There are ways to die that aren't painful, like nitrogen/argon gas or overdose on painkiller medication. These obviously don't provoke pain because they weren't common ways to die in the ancestral environment. (Nitrogen is a painful way to die for burrowing mammals because anoxia was a more common cause of death for them.) Death by fire is much worse than most ways to die not because the evolutionary cost is greater but just because evolution had other reasons to make fire excruciating and no reason to tone down the magnitude of suffering when it was at the level to cause death.
Are happiness and suffering merely relative?
There's a common intuition that emotions are relative to one another rather than absolute. In Winona Laduke's Last Standing Woman (p. 187 and 299) we find this passage:
[The Ojibwe Native Americans] understood that both the making and the unmaking were essential parts of life and necessary to keep the balance. After all, what was dawn without dusk and what was life without constant change? [...] For all the pain and heartache we have felt, there has been and will be, an equal amount of joy. That is how everything works. There is always a struggle to maintain the balance.
In an interview with David Pearce on Singularity 1 on 1, Nikola Danaylov makes a similar observation:
The Buddhists would say it's like left and right, up and down, something or nothing. You have to have one to have the other. You cannot have up without having down. How can you have pleasure without ever having pain?
With levity, we might call this intuition "Newton's third law of emotion": "To every emotion there is always opposed an equal emotion."
This intuition comes from our own experience of hedonic adaptation. It is indeed the case
that we get used to great comfort and complain about even minor inconveniences (think "First World problems") and likewise that people in dire conditions can become habituated to hardship. At the neural level, we see the brain adjust in order to maintain balance. For instance, dopamine spikes due to stimulant use lead to downregulation of dopamine receptors.
But even if we believed in a pure hedonic-setpoint mechanism, this wouldn't mean reducing suffering and increasing happiness were hopeless. As David Pearce notes in reply to Nikola, some people are severely depressed for their entire lives, while others are almost uniformly ebullient. So in the worst case, we could select for individuals with higher hedonic setpoints. More realistically, setpoints are adjustable even within a given person. Many causes of depression are not purely biological, and those are examples of ways in which setpoints can be improved without a corresponding increase of suffering somewhere else.
Moreover, past suffering doesn't always dampen the intensity of future suffering. In some cases, negative experiences can lead to further suffering via depression or other psychological disorders. Traumatic events can lead to posttraumatic stress disorder, which amplifies the intensity of future traumatic events. The opposite of habituation is sensitization, and both of these processes occur in brains in various contexts.
The hedonic treadmill is real and important, and it illustrates why, for instance, consumerism is not a gateway to happiness. But it doesn't show that changing the net amount of suffering in the world is impossible, and indeed the opposite is true.
What is happiness, anyway?
People sometimes suggest that "happiness" is any time your conditions improve relative to where they were. This seems to me an impoverished view of happiness and not one I would endorse. To modify an example from Max Carpendale, suppose you were locked in a single state of enjoyment for a long period. Would we say there was no happiness there?
Of course not, and the reason is that "happiness" describes a richer set of brain operations than a temporal difference of reward status. It refers to a "liking" gloss that's applied in certain brain regions and that generates a conscious thought, "Wow, this feels good." In principle, this could presumably continue for a long time in a relatively constant way, even though in practice, evolution wants to conserve on how much pleasure is applied so as to maintain the fitness-maximizing balance of motivational tradeoffs.
Dopamine does seem to represent changes of expected reward over time, but dopamine is not pleasure. For instance, after a monkey associates a tone with the arrival of juice reward, dopamine begins to spike at the time of the tone rather than at the consumption of the juice. But we know from personal experience that consuming the juice is where the pleasure is experienced.
Suppose someone still maintained that the increase in hedonic levels relative to an unpleasant past was what actually constituted pleasure. In that case, we could construct a mind that had unpleasant memories and felt as though it was just now having much better experiences than a few seconds ago. We could keep resetting this mind every few seconds. It would continually come "up" and never go "down." This alone demonstrates that there's no law of conservation of hedonic value.
Relative vs. absolute suffering brain processes
The "suffering is relative" hypothesis seems to assume that suffering is a matter of comparisons, but I think that's far from obvious. At least the simplest forms of suffering are plausibly the kinds of things that are not relative. For example, if we build an artificial reinforcement-learning agent that gets punished for making a bad move, the simplest thing to do is just to use a constant punishment whenever it makes the bad move. It requires a more complex algorithm to change the punishment based on how much past punishment there has been, or for the agent to transform absolute numerical punishments into an ordered list of how bad some punishments were compared with others.
Extending this idea to more complex agents: It's plausible that some people have many more neural processes that correspond to aversion, pain, fear, stress, etc. than other people do, just like the personal computer of a solitaire aficionado contains many more activations of the solitaire software than the personal computer of someone who has the software installed but rarely uses it.
Insofar as evaluative judgments constitute a component of suffering (e.g., "I wish I were dead"), it's clearly true that some people have those sentiments more often than others over a lifetime, and presumably that correlates with underlying neural processes of feeling bad.
It's plausible that hedonic adaptation and relative judgments are higher-level processes performed on top of the simpler neural processing of raw aversion and pain. But whether this is true is something of an empirical question -- maybe brains make relative comparisons of emotional states very early on in hedonic processing, rather than at a late stage? And people disagree about whether lower-level processes have ethical significance or if only the highest levels of self-reflection do.
Maybe a sophisticated way to interpret the emotional-relativism position is in terms of utility functions. Given a bounded utility function, there's some state that represents the worst possible outcome (normalized utility = 0) and another that represents the best possible outcome (normalized utility = 1). Any bounded utility function can be normalized to the same [0,1] scale, so in that sense, comparisons are relative: Utility functions are unique only up to a positive affine transformation, and only comparisons are possible. There's no unique, absolute number for the utility of something.
But this doesn't yield the conclusion that the emotional relativist seemed to be implying -- that we can't change the total suffering or happiness of the world because emotions are relative. Your utility function has different utility values for different states of the world, so indeed you can make things worse or better. There's no absolute scale for the worth of one state relative to another, but that doesn't matter to your actions; you just do things that make the world better, which includes moving toward states that contain less suffering.
In July 2005 I wrote an essay "Is Happiness Relative? -- A Personal Conjecture" that can be found on pages 28-31 of my old collection of journal entries. It describes more concrete examples where aggregate happiness and suffering seem different on average among various people and animals.
The original discussion of this topic on Felicifia includes replies from others.
Lukas Gloor helped inspire the Second Noble Truth view discussed in this piece.