Experimental Plan¶

Experiment 20230822-IKEM¶

Online: 'https://sake.re/20230822-IKEM'

Repo: 'http://felapton.inrialpes.fr/cakes/20230822-IKEM.git'

Experiment design¶

Agents interact by playing games with objects and adapt their ontologies to agree on decision taking. Intrinsic exploration motivation (individual / social and direct / indirect curiosity, creativity and non-exploration), that influences which object and, if the agent is social, partner(s) are chosen for the interaction, is introduced.

Date: 20230822 (Anaïs Siebers)

Hypothesis: 1) Agents will be able to fulfil their motivation in terms of increased exploration (exploratory motivations) and decreased exploration (non-exploratory motivation). 2) Agents with exploratory motivation will be more complete, but less accurate than the baseline and non-exploration. 3) Curious agents will be more accurate and complete, but converge slower in comparison with creativity. 4) In more complex settings (higher number of agents and properties), exploration leads to more completeness, accuracy, distance and faster convergence than the baseline. 5) Agents with non-exploratory motivation will be less accurate and complete than the baseline. 6) The indirect learning models have a higher accuracy and completeness than the direct models. 7) Agents with socially oriented intrinsic motivation will have less diverse knowledge, but agree more and converge faster than agents with individual intrinsic motivation. 8) Heterogeneously motivated agents will have a higher accuracy and completeness, but lower diversity and converge slower than homogeneously motivated agents.

5 runs; 40000 games

Experimental setting: Agents learn decision trees (transformed into ontologies); choose the objects (and partners) to interact with; adapt by splitting their leaf nodes

Dependent variables: ['number of agents', 'number of properties', 'motivated agents ratio', 'social', 'direct action choice']

Independent variables: ['success rate', 'completeness', 'accuracy', 'ontology distance', 'object exploration', 'partner exploration']

Experiment¶

Date: 20230822 (Anaïs Siebers)

Lazy lavender hash: db26577bee05f00f618e8b97f616ba24356a4afe

Parameter file: params.sh

Executed command (script.sh):

The independent variables have been varied as follows:

Number of Agents: [3,15,30]
Number of Properties: [4,6,8]
Motivated Agents Ratios: # [CURIOUS;CREATIVE;NON-EXPLORATORY]
                        [[0:0:0],[1:0:0],[0:1:0],[0:0:1],[0.5:0.5:0],[0.5:0:0.5],[0:0.5:0.5],[0.3:0.3:0.3]]  
Agent is Social: [true, false]
Direct Action Choice: [true, false]

Raw results¶

Style¶

Data exploration¶

Initialisation & helper function¶

Overview¶

Analysis¶

Full results are available on zenodo.¶

Hypotheses¶

Style¶

Helper functions¶

Proof direct baseline == indirect baseline¶

1) Agents will be able to fulfil their motivation in terms of increased exploration (exploratory motivations) and decreased exploration (non-exploratory motivation).¶

Statistical Analysis¶

Interpretation of the Results¶

The object exploration score is indeed higher for curious ([1:0:0]), creative ([0:1:0]) and even non-exploratory agents ([0:0:1]) compared to the baseline ([0:0:0]). The same applies for the partner exploration score.

One ANOVA was performed, testing only the direct models and baseline. The other ANOVA tested only the indirect models and baseline. Because of the difference of the approach behind the direct and indirect models, they are tested separately. It can be seen, that both the direct models and the indirect models have a statistically significant effect on the object exploration score and the partner exploration score. The post-hoc Tukey HSD test shows, that the influence of the exploration motivation is always significant regarding the baseline scenario.

Non-exploratory agents have the goal to explore less, and therefore the object exploration should be lower than the baseline. This is not the case. But the motivation of curious and creative agents to explore more objects is nevertheless given (see higher object exploration regarding the baseline).

2) Agents with exploratory motivation will be more complete, but less accurate than the baseline and non-exploration.¶

Statistical Analysis¶

Interpretation of the Results¶

It is shown that the baseline plot always converges to a higher knowledge accuracy and completeness than curiosity and creativity. The non-exploration plot always converges to a lower knowledge accuracy and completeness. Regarding non-exploration, exploratory agents have a higher completeness. But, they also have more knowledge accuracy, contrary to the hypothesis. It has to be considered that some results for knowledge accuracy are not statistically significant. Exploratory agents converge to a lower accuracy and completeness than the baseline.

Again, two ANOVA tests were performed, testing the results for the direct and indirect model. ANOVA reveals that there is a statistically significant effect of the different exploration motivations on the completeness and knowledge accuracy of the agents’ knowledge. According to the post-hoc Tukey HSD test, the exploratory motivations (curiosity and creativity) compared with the baseline (non- exploration and baseline) show a significant effect concerning completeness in both (direct and indirect) models. But no statistically significant effect for knowledge accuracy between the exploration motivations and non-exploration can be found. Knowledge accuracy is significant for the exploration motivations regarding the baseline.

3) Curious agents will be more accurate and complete, but converge slower in comparison with creativity.¶

Statistical Analysis¶

Interpretation of the Results¶

Convergence can be observed in the success rate. If the success rate stabilises, the agents do not disagree any more, and thus there are no changes in their knowledge any more. As before, the direct and indirect models are tested separately. It can be seen that curiosity and creativity with the direct models are very similar. ANOVA confirms that no significant effect can be found.

For the indirect models, on the other hand, a statistically significant effect can be found. As hypothesised, curiosity converges more slowly, but to higher knowledge accuracy and completeness than creativity.

4) In more complex settings (higher number of agents and properties), exploration leads to more completeness, accuracy, distance and faster convergence than the baseline.¶

Statistical Analysis¶

Interpretation of the Results¶

The completeness of the baseline always converges to a higher value than curiosity and creativity in the direct and indirect models, except for the simple setting with indirect models. The knowledge accuracy of the baseline is higher in the normal setting than the knowledge accuracy of the direct and indirect models of curiosity and creativity. But in the complex setting, the knowledge accuracy for the baseline is lower than for curiosity and creativity. The ontology distance of the baseline is – except for the simple settings – always lower. Overall, the baseline takes longer to converge for all dependent variables compared to the exploratory motivations. Surprisingly, the ontology distance starts very low in the simple setting and very high in the complex setting. This strongly suggests that there is likely another variable – in this case the numberof properties – strongly influencing and dominating the results. With few properties, the ontologies are very similar whereas a high amount of properties results in very dissimilar ontologies.

The ANOVA shows that there are no significant results for the simple setting with three agents and four properties. For the normal setting (15 agents, six properties), the results of completeness, knowledge accuracy and ontology distance are statistically significant. The results of the success rate are not significant for the direct model. The results of the complex setting (30 agents, eight properties) are significant for completeness, ontology distance and the success rate, but not for knowledge accuracy.

Looking at the post-hoc Tukey HSD for the simple setting, all results are not significant. All dependent variables in the normal setting, except for the success rate, of the direct and indirect models of curiosity and creativity are statistically significant regarding the baseline. In the complex setting, the ANOVA and post-hoc Tukey HSD show that completeness, ontology distance and success rate are significant in the direct and indirect models. But knowledge accuracy is significant for direct creativity.

To summarise, in more complex settings, exploratory motivations (curiosity and creativity) lead to faster convergence. Contrary to the hypothesis, completeness is lower even with increasing complexity, but knowledge accuracy increases with increasing complexity. Exploration always leads to more ontology distance with increasing complexity.

5) Agents with non-exploratory motivation will be less accurate and complete than the baseline.¶

Statistical Analysis¶

Interpretation of the Results¶

Agents with non-exploratory motivation are less accurate and complete regarding the baseline for the direct as well as the indirect model of non-exploration. An ANOVA test confirms the significance of the results and the post-hoc Tukey HSD shows the effect of the motivation (non-exploration or baseline) on completeness as well as accuracy.

6) The indirect learning models have a higher accuracy and completeness than the direct models.¶

Statistical Analysis¶

Interpretation of the Results¶

The figures show the plots of the direct and indirect motivations for completeness and accuracy. The direct and indirect model are compared and not analysed separately. Therefore, they are investigated separately with individual and social agents. The results of the direct and indirect model are always very similar, which has been confirmed by an ANOVA. Only curiosity has an effect on knowledge accuracy and non-exploration on completeness, in the individual case. With social agents, social creativity has an effect on knowledge accuracy and completeness and social non-exploration has an effect on completeness. Otherwise, there are no statistically significant results. For knowledge accuracy in the context of individual curiosity, the indirect model has a higher accuracy. The completeness of individual non-exploration with the indirect model is lower. Social creative and non-exploratory agents have a higher completeness for the indirect model, but social creativity has a lower accuracy for the indirect model.

7) Agents with socially oriented intrinsic motivation will have less diverse knowledge, but agree more and converge faster than agents with individual intrinsic motivation.¶

Statistical Analysis¶

Interpretation of the Results¶

Another aspect addressed by the hypotheses is the effect that socialising agents have on the agents’ knowledge. All figures show that with socially oriented motivation the ontology distance and success rate converge to higher values in less iterations. There is one exception: In the indirect models of non-exploration, the values for socially oriented agents and individual agents converge very similarly. Overall, agents agree more and converge faster when they are socially motivated. Nevertheless, their ontologies are more distanced.

The ANOVA, as well as the post-hoc Tukey HSD show that the effect of whether agents are socially oriented or not is mostly significant. There is no significant effect of socially oriented motivation for direct and indirect non-exploratory motivation.

8) Heterogeneously motivated agents will have a higher accuracy and completeness, but lower diversity and converge slower than homogeneously motivated agents.¶

Statistical Analysis¶