1What the method is for
Asking, through a generative model, how the emotions a place calls up change as the reworking of the environment advances.
Niche construction theory treats the process by which an organism alters its environment and that alteration bears on the selection pressures of the next generation (Odling-Smee, Laland & Feldman 2003). In humans, the use of fire, agriculture, the building of dwellings, the forming of cities and the design of institutions are taken as forms of it. Onohara's research brings emotion into this frame. The question is whether, as the degree of reworking of the environment advances, the pattern of emotions called up in that place changes.
The reason for using a generative adversarial network here is not simply that it can generate. What the method is trying to express is a balance between two pressures. One is the pressure of “does this look adapted to that environment”, carried by the discriminator. The other is the pressure of “is this close to the cultural and emotional features theory holds desirable”, carried by an evaluator supplied from outside. The design is motivated by the question of whether the push and pull between exploration and settling that the theory of cultural evolution speaks of can be written down as an apportionment between these two.
Whether that design motivation — “can it be written down” — captures the mechanism behind the empirical data is a separate matter. Section 7 deals with the latter.
2Origin — two lineages meeting
The computation comes from machine learning, the question from evolutionary biology.
On the computation side: generative adversarial networks
A generator \(G\) and a discriminator \(D\) are trained together; \(G\) aims to make \(D\) err, and \(D\) aims to tell real from fake. The framework was proposed by Goodfellow et al. (2014) and is formulated as the minimax problem of a two-person zero-sum game. The structure in which one side's improvement is the other's deterioration appears directly in the objective function.
On the question side: niche construction theory and cultural evolution theory
The view that organisms alter their environment and that the alteration changes selection pressures was systematized by Odling-Smee et al. (2003). The framework treating culture as a process with inheritance, variation and selection goes back to Boyd & Richerson (1985), and Henrich (2016) argues that cumulative cultural evolution is the key to human success. The process by which the environment for social learning itself grows richer across generations is what Sterelny (2012) described as the “evolved apprentice”. Laland et al. (2015) set out the extended evolutionary synthesis that takes these in.
The push and pull between exploration and objective
What the package side explicitly cites is the argument that pressing towards an objective function can itself obstruct arrival at the objective (Lehman & Stanley 2011). Taking the temporal trade-off between goal-directed optimization and open-ended search and putting it outside the model as a parameter is the idea behind λ-scheduling.
3Correspondence with Perspective 6, games and adversaries
When one side does better, the other is in trouble. Learning is read as that balance.
\(G\) … the generating side; it wants to make fakes that pass for real
\(D\) … the discriminating side; it wants to tell real from fake
\(\min\max\) … each side answers the other's best move with its own best move; the point of balance is the solution
What this research adds
Perspective 6 of Reading the seven perspectives in equations handles this minimax form itself. What Onohara's research adds is that it does not leave the adversarial pressure as the only pressure. A second reward is prepared, for the features theory defines as “desirable”, and the apportionment between the two is moved over time. The next section is the formulation of that.
Perspective 6 also carries distributionally robust optimization, which prepares for the worst case, and the chase between generation and detection in audio deepfakes. All of them have the same structure, in which one side's improvement is the other's deterioration.
The same perspective also carries experiments on cooperative behaviour using economic games. There the players are living people, and the payoffs are not zero-sum. What Perspective 6 handles in common is the condition that each party's choice bears on the others' outcomes, not whether their interests are opposed.
4The formulation of λ-scheduling
This is the core. The apportionment of the two pressures is put outside the model, as a function of time.
\(L_{\mathrm{adv}}\) … the adversarial loss; treated as better when smaller
\(R\) … the theory-driven feature reward; better when larger, which is why it enters with a negative sign
\(\beta\) … the scale of the reward term. Default 0.10
\(\lambda(t)\) … a value in \([0,1]\). Near 0 it favours adversarial learning; near 1 it favours the feature reward
As \(\lambda\) moves from 0 to 1, a temporal switch takes place: early on the weight is on “does this look adapted to that environment”, later on “is this close to the features theory holds desirable”. It can be read as writing the cultural-evolution idiom of exploration giving way to settling as an apportionment inside a loss function.
The three shapes of schedule
scheduler.py sets \(t^{\ast} = \lfloor pT \rfloor\) from the total number of steps \(T\) and the reach portion \(p\), and uses a window width \(w = \lfloor p_w T \rfloor\) to give three shapes.
\(s\) … sharpness. Default 12.0; larger makes the transition steeper
| Parameter | Default | Meaning |
|---|---|---|
total_steps \(T\) | 400 | Total number of steps |
reach_portion \(p\) | 0.30 | Portion of \(T\) by which \(\lambda\) is to reach 1 |
window_portion \(p_w\) | 0.05 | Width of the rise window (delayed / logistic) |
sharpness \(s\) | 12.0 | Sharpness of the logistic |
beta \(\beta\) | 0.10 | Scale of the feature reward |
Defaults of the published package placemotiongan, quoted from the definitions in scheduler.py and losses.py.
Changing the shape of the schedule is the operation this research performs
All three shapes reach \(\lambda=1\) in the end, but the path there differs. analyses/sensitivity.py runs the three conditions under the same random seed and collects the results into one table. What is being measured is whether when the constraint is applied changes the path of learning. Taking the shape along the way, rather than the destination, as the variable operated on is the claim of this package.
5The implementation on the paper's side (WCTP2025)
Survey data from 150 respondents, and generation conditioned on a stage label.
What the paper reports is the collection of empirical data and an analysis by a conditional generative adversarial network built to match it. The foundation is a framework dividing the progress of niche construction into five stages.
- Setting up the stage model Five stages are posited: the natural stage (an untouched environment), the first stage (simple survival strategies), the second stage (learned and routinized behaviour), the third stage (social cooperation) and the fourth stage (planned, intentional modification of the environment). How this framework should be positioned needs care (section 7).
- The survey 150 adults, men and women, aged 20 to 50; conducted on 2 June 2025. The objects of evaluation are the five stages of cultural evolution × 10 places each = 50 places in all.
- Emotional evaluation Each of the eight basic emotions of Plutchik's wheel is given three levels of intensity, making 24 categories. Personality traits are measured on the five Big Five factors using the TIPI-J.
- Defining the features Four cultural features (clarity, cooperation, innovation, transmissibility) and four emotional features (joy response, fear response, trust building, expectation), eight dimensions in all. The evaluator is a weighted sum, set so that the cultural weights total 0.6 and the emotional weights 0.4.
- Conditional generation A 128-dimensional noise vector is concatenated with a 5-dimensional one-hot for the stage, giving a 133-dimensional input to the generator. The generator is a network with residual connections and batch normalization; the discriminators are prepared one per stage, five in all, with spectral normalization applied.
- Training 400 generations, batch size 64. \(\lambda\) is moved from 0.000 to 1.000, with the transition taking 120 generations. All features are z-score normalized and then passed through tanh.
The reported results
On the empirical side, the tendency reported is that as the stage advances fear and disgust are suppressed while trust and joy rise. At the fourth stage expectation stands out and the diversity of emotion (entropy) is at its lowest. The paper calls this “emotional convergence” and positions it as suggesting a new form of stress in a highly managed environment.
| Generated cultural feature | Initial | Final generation | As stated in the paper |
|---|---|---|---|
| clarity | 0.245 | 0.678 | increase |
| cooperation | 0.312 | 0.723 | increase |
| innovation | 0.456 | 0.534 | slight increase |
| fear_response | 0.634 | 0.289 | decrease |
| trust_building | 0.278 | 0.687 | increase |
Quoted as they stand from the values given in section 3.4.2 of the paper, “evolutionary trajectory of cultural features”. These are outputs of the generative model, not observed values from the empirical data.
The discriminator loss figures differ between sources
Section 3.4.3 of the paper's body text gives the discriminator loss by stage (lower is more adapted) as second stage 0.23 (most adapted), fourth stage 0.31, natural stage 0.45 (least adapted). The architecture diagram gan.html that accompanies the same research, however, records first_stage: 0.2511 (best adapted), and the two do not agree on which stage is most adapted.
Which is the final result cannot be determined from the materials at hand. The wording of the paper's body text is taken here, and the fact of the discrepancy is recorded as it stands.
6What the published package provides
This is not the paper's implementation. It is the λ-scheduling framework taken out on its own.
The package's training loop is not training any network
simulate_training in placemotiongan/train.py generates the adversarial loss and the feature reward from formulas for stochastic decay and rise. The adversarial loss starts near 0.80 and decays by 0.35% each step with normal noise added; the feature reward starts near 0.10 and rises with saturation. A comment in the code states plainly that it generates dummy adv_loss and feature_reward with controlled stochasticity.
model.py does define generator and discriminator classes, but they are never called from this training loop. The curves obtained by running the package therefore show how the λ schedule moves the apportionment of the loss; they are not the result of a model having learned anything.
Then what is the published artefact for
The package's own account is consistent. It states that rather than focusing on large-scale neural network training, it emphasizes the explicit modelling of the trade-off between adversarial pressure and the emergence of structured features, through a controllable λ-scheduling mechanism. What it provides is a framework that treats temporal structure as a first-class component, together with the apparatus for observing that behaviour in a deterministically reproducible form.
The package therefore does not reproduce the analysis in the paper. The two share the same idea of λ-scheduling, but the implementations are different things.
The sign of the loss differs between sources
The package's losses.py writes \(L = (1-\lambda)L_{\mathrm{adv}} - \lambda\beta R\), with the reward term subtracted. Since a loss is better when smaller, this is the consistent direction: the larger the reward, the lower the loss. The architecture diagram in gan.html, by contrast, writes Loss = λ×emotion_reward + (1-λ)×adv_loss, with the reward added, and without \(\beta\).
The two cannot both hold as equations, so the package's implementation is quoted here. The diagram may be a simplified notation for exposition, or the implementation on the paper's side may follow a different sign convention. Either way, the only one that can be confirmed from published code is the former.
No run results are published
Neither repository stores run results such as CSV logs or figures. This page therefore gives no output figures from the package side at all. What can be written is the default parameter values and the formulas readable from the code, and nothing more.
7Limits
What the method cannot say. There are four points.
1. The stage model is a hypothetical framework
Of the division into five stages, the paper itself speaks of a “hypothetical five-stage model” and states that the stage model as such is not set out explicitly in prior work. Consistency with existing theory — niche construction theory, cultural evolution theory, the evolutionary theory of cooperative behaviour — is claimed, but this is not a stage division confirmed empirically. Both the generation conditioned on stage labels and the differences between stages obtained in the analysis take on meaning only once this framework is granted.
2. Description bias — the respondents were not looking at places
This is the point section 4.3.1 of the paper raises as a methodological limitation. Analysis with BERT is reported to show that 65% of the joy emotion depends on the text description, and 43% of the fear emotion is affected by the text.
The respondents did not visit the actual places; they read descriptions of places and chose emotions. What was measured is therefore not “the emotion a place calls up” but, at least in part, “the emotion that description calls up”. Whether the differences in emotional pattern between stages come from the degree of reworking of the environment or from the style in which the descriptions were written for each stage is something these data cannot separate. That the paper's future directions include comparative validation against real experience using VR and AR is a response to this limitation.
3. Reproducing a pattern does not mean the mechanism agrees
The paper states that the generated emotional patterns showed high agreement with the empirical data. But a generative model's being able to reproduce an observed pattern does not mean that its generative process is the same as the mechanism of real cultural evolution. A model with enough freedom can produce the same distribution from a different mechanism. What has been confirmed here is an existence claim — that under this formulation it is possible to produce this pattern — not the claim that cultural evolution actually proceeds by this mechanism.
Still less so, given that the feature weights of the evaluator are values set by the analyst. Once what counts as a “desirable cultural feature” has been decided in advance and a reward given that moves the output in that direction, the output moving in that direction is not a result but a design.
4. What is published and what is reported are different
As section 6 sets out, the published package is a simulation and does not reproduce the paper's experiment. The implementation code on the paper's side — the per-stage discriminators, the residual generator, the evaluator — is not included in the published repositories. No means is currently provided for a third party to verify the reported figures. That a package flying the flag of reproducibility stands alongside it makes this all the easier to misread.
8Sources and data
The paper is collected in the proceedings of WCTP 2025. The software is published in two repositories, of which cultural-dynamics-gan is the version organized as a package.
- Onohara, A. & Ouchi, H. (2026) GAN-Based Modeling of Emotional Dynamics in Cultural Evolution and Niche Construction: An Integrated Empirical Approach. Proceedings of the Workshop on Computation: Theory and Practice (WCTP 2025), Atlantis Press. DOI: 10.2991/978-94-6239-638-8_25
- Onohara, A. “PlaceEmotion-GAN” github.com/aonoa68/cultural-dynamics-gan (package version, MIT License)
- Onohara, A. “PlaceEmotion-GAN” github.com/aonoa68/PlaceEmotionGAN DOI: 10.5281/zenodo.17401003
- Survey instrument: niche_construction_survey (survey items measuring the environmental impact of niche construction and cooperative behaviour)
- JSPS KAKENHI Grant-in-Aid for Early-Career Scientists, “Elucidating the emotional relationship between speech and place using an evolutionary approach” (Grant Number JP24K16071, FY2024–FY2026)
References for the method
- Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. & Bengio, Y. (2014) Generative Adversarial Nets. Advances in Neural Information Processing Systems, 27, 2672–2680.
- Lehman, J. & Stanley, K. O. (2011) Abandoning Objectives: Evolution Through the Search for Novelty Alone. Evolutionary Computation, 19(2), 189–223. doi:10.1162/EVCO_a_00025
- Oshio, A., Abe, S. & Cutrone, P. (2012) Development, reliability and validity of the Japanese version of the Ten Item Personality Inventory (TIPI-J). Japanese Journal of Personality, 21(1), 40–52.
References for the theory
- Odling-Smee, F. J., Laland, K. N. & Feldman, M. W. (2003) Niche Construction: The Neglected Process in Evolution. Princeton University Press.
- Boyd, R. & Richerson, P. J. (1985) Culture and the Evolutionary Process. University of Chicago Press.
- Sterelny, K. (2012) The Evolved Apprentice: How Evolution Made Humans Unique. MIT Press.
- Laland, K. N., Uller, T., Feldman, M. W., Sterelny, K., Müller, G. B., Moczek, A., Jablonka, E. & Odling-Smee, J. (2015) The extended evolutionary synthesis: its structure, assumptions and predictions. Proceedings of the Royal Society B, 282(1813), 20151019. doi:10.1098/rspb.2015.1019
- Henrich, J. (2016) The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter. Princeton University Press.
The paper's reference list carries Tomasello (2014) A Natural History of Human Cooperation, MIT Press, but the bibliographic record for a work of that title could not be confirmed. The 2014 work by that author is A Natural History of Human Thinking (Harvard University Press). Since it could not be confirmed, it is not listed here.
Which perspective this method belongs to
A generative adversarial network formulates learning as the balance between two parties in competition. When the generator improves, the discriminator is in trouble; when the discriminator improves, the generator is. The solution is the point at which neither can improve unilaterally. Perspective 6 (games and adversaries) in Reading the seven perspectives in equations applies directly.
This research, however, mixes a theory-driven reward into the adversarial pressure and moves the apportionment over time. It is not a pure minimax but a composition with an objective function supplied from outside the game, and that is where it departs from the textbook setting.
This page sets out, for readers in the same field, a method Onohara uses in actual research. Only values recorded in the paper and in the published code are quoted, and which source each came from is stated in the text. No estimates or approximations are given.
The whole layout of where the methods sit is in Reading the seven perspectives in equations; which teaching session falls under which perspective is in Methods & Mathematics.