Pills from ASP Challenge 2019 Mario Alviano Department of Mathematics and Computer Science University of Calabria, Italy, EU LPNMR 2019 1 / 17
Overview 1 Fastfood Problem Two lines encoding Use median if only one deposit (another two lines) 2 / 17
Overview 1 Fastfood Problem Two lines encoding Use median if only one deposit (another two lines) 2 House Reconfiguration Problem Around 80 lines of ASP , 16 of which where weak constraints Only toy instances available 3 Insurance Referees Assignment Problem Around 50 lines of ASP Only toy instances available 2 / 17
Overview 1 Fastfood Problem Two lines encoding Use median if only one deposit (another two lines) 2 House Reconfiguration Problem Around 80 lines of ASP , 16 of which where weak constraints Only toy instances available 3 Insurance Referees Assignment Problem Around 50 lines of ASP Only toy instances available 4 Checking Policies for Reactive Agents Input, output, desired function unclear to me 2 / 17
Overview 1 Fastfood Problem Two lines encoding Use median if only one deposit (another two lines) 2 House Reconfiguration Problem Around 80 lines of ASP , 16 of which where weak constraints Only toy instances available 3 Insurance Referees Assignment Problem Around 50 lines of ASP Only toy instances available 4 Checking Policies for Reactive Agents Input, output, desired function unclear to me 5 Automated Warehouse Scenario Let’s talk more about this one! 2 / 17
Automated Warehouse Scenario (1) Robots moving in a grid and under shelves Shelves cannot be put down on highway cells Goal Move shelves to picking stations and deliver products 3 / 17
Automated Warehouse Scenario (2) Theorem (Peter Schuller, 2013 — private communication) Robots involved = ⇒ video 4 / 17
Example instances from 2019 Small size 4x4 grid 2 robots around 4 delivers 5 / 17
Example instances from 2019 Small size 4x4 grid 2 robots around 4 delivers Good for... Understanding the problem Finding bugs (maybe) Let’s check one of them 5 / 17
Example instances from 2019 Small size 4x4 grid 2 robots around 4 delivers Good for... Understanding the problem Finding bugs (maybe) Let’s check one of them Bad for... almost everything else! Just bruteforce for the solution No hint on the difficulty of realistic instances 5 / 17
Example instances from 2017 Medium size at least 12x12 grids at least 6 robots tens of delivers Bigger instances from the ASPRILO project https://asprilo.github.io 6 / 17
Example instances from 2017 Medium size at least 12x12 grids at least 6 robots tens of delivers Bigger instances from the ASPRILO project https://asprilo.github.io Gave up in 2017 Pure (poor) ASP restriction I had an ASP encoding, but worked only on the 4x4 toy instance in the description of the problem 6 / 17
Is ASP bad or good? 7 / 17
Is ASP bad or good? ASP by itself is bad! We use SQL to query relational databases We don’t use SQL to build websites and apps We combine SQL with other programming languages 7 / 17
Is ASP bad or good? ASP by itself is bad! We use SQL to query relational databases We don’t use SQL to build websites and apps We combine SQL with other programming languages ASP itself is good! We have to use ASP for what it is good for We have to use ASP to solve subtasks of a bigger problem We have to combine ASP with other languages 7 / 17
Is ASP bad or good? ASP by itself is bad! We use SQL to query relational databases We don’t use SQL to build websites and apps We combine SQL with other programming languages ASP itself is good! We have to use ASP for what it is good for We have to use ASP to solve subtasks of a bigger problem We have to combine ASP with other languages ASP Challenge 2019 gave sufficient freedom for this! 7 / 17
Classical planning approaches Uninformed search Breadth first Depth first Iterative deepening Informed search Dijkstra’s algorithm A ∗ Best first 8 / 17
Classical planning approaches Uninformed search Breadth first Depth first Iterative deepening Too many plans to test Informed search Dijkstra’s algorithm A ∗ Best first Require a priority queue to store (too many) states 8 / 17
What’s wrong with them? Optimality is expensive 9 / 17
What’s wrong with them? Optimality is expensive Can we at least have some plan with ASP? 9 / 17
What’s wrong with them? Optimality is expensive Can we at least have some plan with ASP? 9 / 17
My solution: Python controller using ASP modules [everything delivered] [else] matcher distancemeter planner STOP 10 / 17
My solution: Python controller using ASP modules [everything delivered] [else] matcher assign shelves to picking stations distancemeter planner STOP 10 / 17
My solution: Python controller using ASP modules [everything delivered] [else] matcher assign shelves to picking stations distancemeter compute distances from targets (shelves, picking stations, put-down cells) planner STOP 10 / 17
My solution: Python controller using ASP modules [everything delivered] [else] matcher assign shelves to picking stations distancemeter compute distances from targets (shelves, picking stations, put-down cells) planner robots, do something! try to deliver, or try to reach a target shelf, or STOP try to get close to a target shelf 10 / 17
Pills of my solution matcher.lp — around 60 lines of ASP code Use ordered encoding 11 / 17
Pills of my solution matcher.lp — around 60 lines of ASP code Use ordered encoding distancemeter.lp — around 60 lines of ASP code Plus 10 lines of template Distance in the predicate to avoid recursive negation 11 / 17
Pills of my solution matcher.lp — around 60 lines of ASP code Use ordered encoding distancemeter.lp — around 60 lines of ASP code Plus 10 lines of template Distance in the predicate to avoid recursive negation planner.lp — around 150 lines of ASP code Only state changes, no actions! What is the knowledge encoded by pick up at time 100? 11 / 17
Pills of my solution matcher.lp — around 60 lines of ASP code Use ordered encoding distancemeter.lp — around 60 lines of ASP code Plus 10 lines of template Distance in the predicate to avoid recursive negation planner.lp — around 150 lines of ASP code Only state changes, no actions! What is the knowledge encoded by pick up at time 100? controller.py — around 400 lines of python code Interpret states and inject actions Anytime visualization of the plan 11 / 17
Pills of matcher.lp 12 / 17
Pills of distancemeter.lp 13 / 17
Pills of planner.lp (1) 14 / 17
Pills of planner.lp (2) 15 / 17
Let’s run it! The plan is produced a few steps at time Never look at the past 16 / 17
Let’s run it! The plan is produced a few steps at time Never look at the past Don’t even think to optimality In a real environment, after a few actions there will be a mistake and you have to replan anyhow 16 / 17
Let’s run it! The plan is produced a few steps at time Never look at the past Don’t even think to optimality In a real environment, after a few actions there will be a mistake and you have to replan anyhow Potentially, one could add new orders during the computation of the plan 16 / 17
Conclusion Optimality is a chimera here Having some plan is also a chimera for ASP by itself 17 / 17
Conclusion Optimality is a chimera here Having some plan is also a chimera for ASP by itself Plans can still be computed by taking advantage of ASP Apparently not too far from optimality (tested on very small instances) 17 / 17
Conclusion Optimality is a chimera here Having some plan is also a chimera for ASP by itself Plans can still be computed by taking advantage of ASP Apparently not too far from optimality (tested on very small instances) 17 / 17
Recommend
More recommend
