Rationale

The ~4000 data sets discussed here represent the choice behavior of 4000 virtual agents animated by ETBD (McDowell, 2004). They were generated between 2010 - 2013 (~100 days of CPU time) as part of my dissertation project (Popa, 2013). 

About 600 (~100 conditions) were discussed in the thesis (Phase 1, pp. 15 - 50) and 360 of these 600 (72 conditions) ​were discussed in Popa & McDowell (2016).  

All data - 25 GB spread across 800 .mdb files - was reanalyzed in 2019. Out of 4000 datasets, 56 had missing values and were discarded.  The remaining 3946 (the summaries) were scaled (z-scores*) and consolidated in one dataset, with each row representing behavioral descriptors for one digital agent. 

 

The purpose of this summary is to test machine learning algorithms. This is a particularly challenging task because the target variables describe emergent properties of complex systems, and the evolution of complex systems is governed by local, non-deterministic rules. In this case, the systems are populations of phenotypes, pushed through time by Darwinian rules implemented in ETBD. This work is ongoing; details can be shared upon request. 

Design and measures

 

(for details, see Popa, 2013, pp. 16-23, and 23-27)

 

Independent variables: 

  1. reinforcement rate (5 values)

  2. reinforcement magnitude (5 values)

  3. Hamming COD (4 values)

  4. mutation rate (8 values)

The fully crossed design entailed 800 experimental conditions. Five agents were ran in each condition. Each agent experienced a concurrent-schedule environment with 11 pairs of Variable Interval schedules (VI VI), of positive reinforcement, each in effect for 20000 generations.

Behavioral measures:

  • sensitivity to reinforcement (a

  • frequency target behavior (b1 + b2)

  • frequency of acquired reinforcers (r1 + r2)

  • frequerncy of changeovers (CO)

  • frequency of bouts

  • average bout length (in responses)

  • average bout duration (in generations; NEW)

  • average Inter-Bout Interval (IBI)

  • sustained behavior (%) = all target behavior/bout behavior

  • topographic variability (average distance between consecutive emissions)

Functional and conceptual equivalence between the digital and the real world. The first two panels show concurrent schedules environments for human and digital agents, respectively. Behavior exhibited in these environments can be analyzed in the same way, as illustrated in the third panel.The two response sequences are identical. Each entails 12 behaviors (b1 + b2), 4 changeovers, and 2 bouts of 5 and 3 responses, respectively. Bout, or sustained behavior, made about 66.7 of all target behavior (8/12). Topographic variability is given by  (5 + 37 +  ... + 0 + 0) / 11. 

Student

POINTS

22

ETBD

  ...    471 ... 511    512 ... 552   ...   4095 

Target Class 1

Target Class 2

time

ETBD

Student

t1

t2

t3

t4

t5

t6

t7

t8

t9

t10

t11

t12

516

511

548

533

500

480

475

476

499

522

522

522

B2

B1

B2

B2

B1

B1

B1

B1

B1

B2

B2

B2

Copyright 2019 Andrei Popa