Decision-Theoretic, High-Level Agent Programming in the Situation Calculus Export

@inproceedings{boutilier00decisiontheoretic, abstract = {We propose a framework for robot programming which allows the seamless integration of explicit agent programming with decision-theoretic planning. Specifically, the DTGolog model allows one to partially specify a control program in a highlevel, logical language, and provides an interpreter that, given a logical axiomatization of a domain, will determine the optimal completion of that program (viewed as a Markov decision process). We demonstrate the utility of this model with results...}, address = {Menlo Park, CA, USA}, author = {Boutilier, Craig and Reiter, Raymond and Soutchanski, Mikhail and Thrun, Sebastian}, booktitle = {Artificial Intelligence (AAAI-00) and of the 12th Conference on Innovative Applications of Artificial Intelligence (IAAI-00)}, citeulike-article-id = {391363}, citeulike-linkout-0 = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.5.6867}, comment = {The authors present DTGolog, a decision theoretic extension of the agent programming language Golog. DTGolog deals with uncertainty and general reward functions, and similarly to Golog it can be viewed as a language and methodology with which to provide ``advice'' to a decision-theoretic planner. The user provides a background action theory as well as an optimization theory which defines the reward function and some optimality criterion. Apart from normal deterministic actions, primitive actions can be stochastic. Each stochastic action is associated with a list of deterministic actions and a probability with which each one can happen. This is done using the predicates \$stochastic(a,s,n)\$ and \$prob(n,p,s)\$ for a stochastic action \$a\$ and the possible nature's action \$n\$. The semantics of FTGolog is specified in a similar way as in Golog, with the predicate \$BestDo\$ playing the role of \$Do\$. \$BestDo(prog,s,horz,pol,val,prob)\$ computes a policy, an expanded Golog program in which every nondeterministic choice point is grounded with the selection of an optimal choice. The structure of a policy is a conditional plan based on sensing actions which sence the nature's choice for each stochastic action. The authors show how DTGolog is implemented in Prolog and discuss results from applying this fraework to an RWI B21 robot for a simple application.}, keywords = {ai, cogrobo, depth, golog, logic, mdp, reasoning\_about\_action, robot, situation\_calculus, system}, month = {July}, pages = {355--362}, posted-at = {2005-11-13 03:15:17}, priority = {0}, publisher = {AAAI Press}, title = {Decision-Theoretic, High-Level Agent Programming in the Situation Calculus}, url = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.5.6867}, year = {2000} }

TY - CONF ID - boutilier00decisiontheoretic L3 - citeulike-article-id:391363 N2 - We propose a framework for robot programming which allows the seamless integration of explicit agent programming with decision-theoretic planning. Specifically, the DTGolog model allows one to partially specify a control program in a highlevel, logical language, and provides an interpreter that, given a logical axiomatization of a domain, will determine the optimal completion of that program (viewed as a Markov decision process). We demonstrate the utility of this model with results... AD - Menlo Park, CA, USA EP - 362 TI - Decision-Theoretic, High-Level Agent Programming in the Situation Calculus PB - AAAI Press T2 - Artificial Intelligence (AAAI-00) and of the 12th Conference on Innovative Applications of Artificial Intelligence (IAAI-00) SP - 355 KW - ai KW - cogrobo KW - depth KW - golog KW - logic KW - mdp KW - reasoning_about_action KW - robot KW - situation_calculus KW - system N1 - The authors present DTGolog, a decision theoretic extension of the agent programming language Golog. DTGolog deals with uncertainty and general reward functions, and similarly to Golog it can be viewed as a language and methodology with which to provide ``advice'' to a decision-theoretic planner. The user provides a background action theory as well as an optimization theory which defines the reward function and some optimality criterion. Apart from normal deterministic actions, primitive actions can be stochastic. Each stochastic action is associated with a list of deterministic actions and a probability with which each one can happen. This is done using the predicates $stochastic(a,s,n)$ and $prob(n,p,s)$ for a stochastic action $a$ and the possible nature's action $n$. The semantics of FTGolog is specified in a similar way as in Golog, with the predicate $BestDo$ playing the role of $Do$. $BestDo(prog,s,horz,pol,val,prob)$ computes a policy, an expanded Golog program in which every nondeterministic choice point is grounded with the selection of an optimal choice. The structure of a policy is a conditional plan based on sensing actions which sence the nature's choice for each stochastic action. The authors show how DTGolog is implemented in Prolog and discuss results from applying this fraework to an RWI B21 robot for a simple application. AU - Boutilier, Craig AU - Reiter, Raymond AU - Soutchanski, Mikhail AU - Thrun, Sebastian PY - 2000/07// UR - http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.5.6867 ER -