ACT-R models of training

Transcription

1 ACT-R mdels f training Cletilde Gnzalez Brad Best, Varun Dutt, Octavi Juarez Dynamic Decisin Making Labratry (DDMLab) Carnegie Melln University

2 Agenda A cgnitive accunt f fatigue in a data entry task Summary f theretical insights (paper in preparatin) The Training difficulty principle in the RADAR task The RADAR task: A visual, auditry, memry and decisinmaking cmpnents ACT-R mdels f training invlving unrelated difficulty ACT-R mdels f training under cnditins f related difficulty Summary f this year s accmplishments and next steps DDMLab - September 7,

3 The 2x2 levels f ACT-R (Andersn & Lebiere, 1998) Symblic SubSymblic Declarative Memry Chunks: declarative facts Activatin f chunks (likelihd f retrieval) Prcedural Memry Prductins: If (cnd) Then (actin) Cnflict Reslutin (likelihd f use) Likelihd f retrieval Activatin Base level learning Latency Cnflict Reslutin Utility Prbability learning A i = B i + # j " j B i = ln t j!d T i = F! e " A i W j! S ji + " A U i = P i!g " C i +# U Succ P i = i Succ i + Fail i DDMLab September 27,

4 Speed Accuracy trade ff with extended practice in data entry Frm Healy et al., 2004 DDMLab - September 7,

5 ACT-R 5.0 mdel f fatigue in Data Entry Task Initiatin time Executin time Cnclusin time Encde next number Retrieve key lcatin Type next number Hit Enter All numbers encded All numbers typed Speedup: Prductin cmpilatin Frm Visual Retrieval (key lc) Mtr T Visual Mtr faster access t key lcatin Decrease in Accuracy: A gradual decrease in surce activatin (W) A i = B i + W j! S ji " " + MP k! Sim kl + N(0, s) j Gradual decrease in gal value (G) U = P # G " C +! i i k i U

6 An accunt f fatigue in ACT-R Cgnitin* Affect Reduce accuracy and respnse time in extended, repetitive tasks Cgnitive factrs: Maintenance f gals, Priritizatin f gals, Attentin t nvel versus familiar infrmatin Emtinal factrs: Willingness t engage in an effrtful task (Mtivatin) If, ur thery is crrect, best fits t human data wuld ccur nly when BOTH, W and G decay with practice. DDMLab - September 7,

7 Data cllected under this prject include: One Mdel f the Data Entry task fits data frm 3 experiments: Experiment 1: The speed-accuracy tradeff (Healy, Kle, Buck- Gengler and Burne, 2004) Experiment 2: Anther prf f speed-accuracy tradeff (Healy, Kle, Buck-Gengler and Burne, 2004) Experiment 3: Cgnitive and mtric stressrs (Kle, Healy and Burne, in press) The same mdel results in predictins that supprt the Cgnitive-Affective interactin accunt f fatigue DDMLab - September 7,

8 The W*G Predictins Experiment 1 DDMLab - September 7,

9 The W*G Predictins Experiment 2 DDMLab - September 7,

10 Summary f mdel predictins DDMLab - September 7,

11 In Cnclusin: Extended practice in repetitive tasks results in a speed-accuracy tradeff Accrding t ur ACT-R mdels reduced reactin time is due t prductin cmpilatin while increased errrs are due t a fatigue effect: W*G Best fit t human data is fund when bth, W and G decay ver time Neither f the W r G parameters alne prduce the fatigue effect, supprting the Cgnitin*Affect hypthesis DDMLab - September 7,

12 Agenda A cgnitive accunt f fatigue in a data entry task Summary f theretical insights (paper in preparatin) The Training difficulty principle in the RADAR task The RADAR task: A visual, auditry, memry and decisinmaking tl ACT-R mdels f training invlving unrelated difficulty ACT-R mdels f training under cnditins f related difficulty Summary f this year s accmplishments and next steps DDMLab - September 7,

13 The RADAR task (frm Gnzalez & Thmas, 2007) Radar Grid Threat Sensrs Weapn System Sensrs F D Quiet Airspace Reprt Target Set (Memry Set) G J Guns Ignre Missiles Ttal Blck Scre DDMLab - September 7,

14 A vide f the RADAR task DDMLab - September 7,

15 ACT-R 6.0 mdels f the training difficulty principle Strategy-based (prductin-based) mdels f RADAR invlve: Cgnitive Mdeling f Visual Search Cgnitive Mdeling f parallel auditry prcessing Cgnitive mdeling f Decisin Making The ACT-R mdels reprduce data frm 2 experiments DDMLab - September 7,

16 DDMLab - September 7,

17 ACT-R 6 update Reinfrcement amng strategies: Utility and reward If U i (n-1) is the utility f a prductin i, after its n-1st applicatin and R i (n) is the reward the prductin receives fr its nth applicatin, then its utility U i (n) after its nth applicatin will be U ( n) = U ( n! 1) + [ R ( n)! U ( n! 1) i i " Ui / s e P i = U j / e! j i s i DDMLab - September 7,

18 Visual Search Strategies: the gd, the bad, and the ugly Main Learning in the mdel ccurs thrugh the selectin amng visual search strategies, determined by the utility f the prductins Fur strategies (4 prductins) were identified: The gd: Exhaustive cmparisn (high hits, lw fa) The bad: Randm cmparisn (lw hit, high fa) The ugly: Partial fund (high hits, high fa) The ugly: Partial cmparisn (lw hits, lw fa) DDMLab - September 7,

19 Training difficulty principle: Experiment 1 Main results frm Healy s experiment: CM/VM effects and lad effects Training difficulty did nt help in this task: tne cunting during training hurts test perfrmance DDMLab - September 7,

20 Experiment 1: Effects f mapping & lad at test DDMLab - September 7,

21 Experiment 1: Effects f tne when testing in CM/VM Detectin time Detectin Accuracy (d ) DDMLab - September 7,

22 Experiment 1: Effect f tne when testing with Tne/Ntne DDMLab - September 7,

23 Training difficulty principle: Experiment 2 Decisin making task, related t visual detectin (VM, high lad nly, and detectin f planes amng planes) Main results frm Healy s experiments: Damaging effect f Decisin Making task High false alarms with DM Lw Hits with DM Lw d with DM DDMLab - September 7,

24 Experiment 2: high FA rates with decisin making task DDMLab - September 7,

25 Experiment 2: Effects f decisin making Main effect f DM Interactin f DM task and sessin DDMLab - September 7,

26 Understanding Reinfrcement Learning Amng Strategies DDMLab - September 7,

27 Summary f this year s accmplishments Cncluded a theretical view f fatigue thrugh mdeling in ACT-R (paper in preparatin): Reduced accuracy in extended, repeated tasks is explained by the interactin f Cgnitive and Affective states Created ACT-R 6.0 mdels fr RADAR: Strategy-based implementatin Use f new utility mechanism Reprduced the effects f the Training difficulty experiments 1 & 2 (experiment 2 wrk is in prgress) DDMLab - September 7,

28 Plans fr next year Cnclude the cgnitive view f Fatigue paper Cnclude a paper n the RADAR task including the behaviral and mdel data fits Perfrm an investigatin f the Instance- Strategy mdeling appraches in the RADAR task Start mdeling n Prctr's findings n the S-R cmpatibility DDMLab - September 7,