dataResult <-processDataTest1(path, prefix, names)

Analysis of all learning parameters. Runs with 1000 ticks

With these constant variables: munsell, radius, 9, chordal, tole 0.2

In this analysis we are determining which exploration factor would be the best option to run the experiments

dataResult <-processDataTest2(path, prefix, names)

Charts

Observation

Overall, the exploration factors 20 and 40 achive quality stability earlier than others.

Analysis of all learning parameters. By Decreasing Factor. Exploration rate set to 40

With these constant variables: munsell, radius, 9, chordal, tole 0.2

In this analysis we are determining which decreasing factor would be the best option to run the experiments. We ran 10 times the simulation with each one of the decreasing factor values: 20, 40, 60, 80 100.

dataResult <- processDataTest3()
## [1] "data/Test_03_ByDecreasingFactor_Radius12_1000Ticks_ L_Rate1_Exploration40/munsell, radius, 12, chordal, tole 0.2, lRate 1, decr 20, expl 40, run 6.csv.txt"
## [1] 2
## [1] "Empty File"

## Warning: Removed 961 row(s) containing missing values (geom_path).

Observation

Overall, the decreasing factor 80 achive quality stability earlier than others.

By Average

Analysis of decreasing factor slopes

## Warning: `data_frame()` is deprecated as of tibble 1.1.0.
## Please use `tibble()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_warnings()` to see where this warning was generated.

Analysis of Tolerance and Radius of Interaction Domain on Learning Quality

Once we calibrated the learning parameters of the learning agent (Decreasing Factor 80, Exploration 40, Learning rate 1), we seek to understand to what extent the quality of learning is affected by the number of agents it interacts with and the general tolerance to imperfect action. Our intuition is that the quality of learning is negatively effected by the tolerance because the agent’s assessment of its model highly depends on the responsive actions of every the surrounding agent. We also expect that the quality of learning benefits from a greater number of agents because the agent’s model is informed by a larger number of perceived hues.

6X6 factorial experiment design

The study is designed as a 6X6 factorial experiment.The independent variables or factors of the experiment are i) the radius of the interaction and ii) the tolerance to imperfection. The radius of interaction determines the chance of an agent to interact with others. It has six levels with values 10,12,14,16,18,20, where lower values represent lower chances of interaction. Usually, the radius at value 10 encompasses 5 agents, whereas at value 20 encompasses all the 20 agents present in the simulation. The tolerance to imperfection factor defines whether or not an agent performs an action given the magnitude of the discrepancy between the current condition of the world and the desired condition of the world. If the agent has low tolerance to discrepancy it will adjust its location more often. This factor has six levels which values are 0, 0.2, 0.4, 0.6, 0.8, 1, where lower values represent higher changes of performing a move. At zero tolerance the agent is extremely sensitive to imperfection and it is constantly doing micro-adjustments.

The dependent variables observed in the study are i) the global social viscosity, ii) the quality perceived by the agent, named relative learning quality, and iii) the similarity between the agent’s color model and the dominant model used by every other agent, named compared learning quality. Daniela, please define relative learning quality and compared learning quality It needs to be clear what are we measuring here.

Data processing

For each pair of conditions we ran the simulation 10 times over 1000 ticks. Since we have ten runs for each combined independent conditions, we are looking for a mechanism to summarize them into a single dataset preserving the highest amount of information. For this experiment report we did a simple average of q values in each set of ten runs.

A better mechanism would be to derive a Markov chain of quality probabilities from the data in the pool of datasets. So, the result would be a sequence of probabilities that account for what are the chances of having a quality level given the precedent qualities, that in turn responded to the contextual circumstances in which the agent learned. But we are not ready to make this analysis at this point. We will keep on looking for the right way to do this solution. We anticipate we might need to discretize the quality values to assign probabilities to each q value.

We would also like to understand what is the effect of the learning agent on the social viscosity.To do this, it is necessary to compare the viscosity of a social group with and without learning agent.

## $stat_sig
##   10 12 14 16 18 20
## 1  *  *     *  *  *
## 2  *  *     *     *
## 3  *  *  *  *     *
## 4  *  *  *  *  *   
## 5  *  *  *  *  *   
## 6  *  *  *  *      
## 
## $delta_vis
##            10          12          14          16          18           20
## 1  0.07209351  0.08834549  0.01410811  0.07702661  0.09217726  0.032944236
## 2  0.07984981  0.04737404  0.01076806  0.06574028  0.01302385  0.036281556
## 3 -0.08653710  0.01124390 -0.08422348  0.09434301  0.01524096  0.066941648
## 4 -0.05277619 -0.18015831 -0.24213606  0.06028018 -0.13716954  0.009547290
## 5 -0.43543368  0.12068066 -0.05375751 -0.18653334  0.04117949 -0.007089287
## 6  0.29208584 -0.14688476 -0.15748652 -0.10750520 -0.01298426 -0.022792103

Results on social viscosity

In general terms we observe that interaction radius and the tolerance to imperfection have a positive correlation with the global viscosity. In terms of the combined effect of these two factors, the effect of tolerance appears stronger at higher radius levels, but this observation still needs to be confirmed. On average, the distribution of data is exponential.

The area under the curve accounts for the aggregated viscosity experienced by the whole group of agents in each combined condition over the duration of the simulation. Across tolerance values, we observe that the higher the tolerance the higher the amount of aggregated viscosity. Overall, the radius has a negative effect on the aggregated viscosity except for tolerance condition equals to 0 (no-tolerance) where the effect is positive. A comparison between the AUC at radius 10 and radius 20 shows that the strength of the tolerance effect on social viscosity is attenuated in higher radius levels. This means that the larger the number of interactants, the higher the acceptance of imperfection.

Results on q-value or expected return from starting from state ‘s’ at time ‘t’, taking action ‘a’

Results on compared learning quality