You are on page 1of 9

An Evolutionary Approach to Intelligent Planning

Shikha Mehta1, Bhuvan Sachdeva2, Rohit Bhargava3, and Hema Banati1


1

Department of Computer Science, Dyal Singh College, University of Delhi, Delhi 2 Movico technologies, Gurgaon 3 Accenture , Mumbai {mehtshikha,bhuv.sac,bhargavarohit15}@gmail.com, banatihema@hotmail.com

Abstract. With the explosion of information on WWW, planning and decision making has become a tedious task. The huge volume of distributed and heterogeneous information resources and the complexity involved in their coordination and scheduling leads to difficulties in the conception of optimal plans. This paper presents an intelligent planner which uses modified Genetic Algorithm assisted Case Based Reasoning (CBR) to solve the cold start problem faced by CBR systems and generates novel plans. This approach minimizes the need of populating preliminary cases in the CBR systems. The system is capable of generating synchronized optimal plans within the specified constraints. The effectiveness of the approach is demonstrated with the help of case study on e-Travel Planning. Rigorous experiments were performed to generate synchronized plans with one hop and two hops between train and flight modes of transport. Results proved that GA assisted CBR outperforms the traditional CBR significantly in providing the number of optimized plans and solving cold start problem. Keywords: CBR-GA hybrid, Intelligent Planner, GA based cost optimization, Cold Start Problem, Synchronized Travel Plans.

1 Introduction
Since times immemorial, planning is ubiquitous in everyday life-from meal planning to Day Planning, Activity Planning, Travel Planning, Investment Planning, Tax Planning, Project Planning etc. With the world getting closer, man has more options to choose. However searching, retrieving and analyzing information from heterogeneous and distributed information resources for day-today planning is a wearisome task. This has motivated the researchers to explore the possibilities of automating the planning task using Artificial Intelligence techniques. One of the most successful AI technology applied in commercial and industrial applications is Case Based Reasoning (CBR).CBR [4] is a problem solving paradigm that emphasizes reasoning and planning from prior experience. Each case in CBR based systems may represent a plan or sub-plan to solve a particular problem. CBR retains every new experience in the memory, each time a problem is solved, making it immediately available for future problems. However CBR proves inefficient in a situation for
B.K. Panigrahi et al. (Eds.): SEMCCO 2010, LNCS 6466, pp. 305313, 2010. Springer-Verlag Berlin Heidelberg 2010

306

S. Mehta et al.

which it has no past experience like system proposed by [1] [3] [10] may fail to provide any suggestions to the user if the initial cases are not populated manually. This problem is popularly known as Cold-Start problem and is more prevalent in the systems which depend upon past solutions to solve future queries. Thus various approaches have been integrated with the CBR to analyze the best algorithm or technique to improve the performance of CBR like Meta reasoning tried to improve the efficiency of a cased based reasoning approach but on few occasions it introduced unwanted changes that degraded the system performance [9]. In the current scenario integration of evolutionary approach like Genetic Algorithm (GA) with CBR has been observed as successful innovation over traditional CBR technique. Genetic algorithms (GAs) are a population-based Meta heuristics invented by John Holland in the 1960s [5].Genetic algorithms are evolutionary search and optimization algorithms based on the mechanics of natural genetics and natural selection. They are a family of computational models that use the notion of evolution to generate successively better solutions from previous generations of solutions. The hybrid (CBR-GA) model named SOCBR (Simultaneous Optimization of CBR) outperformed the other models namely COCBR (Conventional CBR), FWCBR (CBR with Feature Weighting), ISCBR (CBR with Instance Selection), and SOCBR (Simultaneous Optimization of CBR) in judging the status of bankrupt or non-bankrupt [7]. Besides the application of GA has also been exploited in different phases of CBR like CBR-GA hybrid for Case retrieval was demonstrated in travel and foodmart [3], tablet formulation problem [8] and Corporate Bond Rating [11]. The utility of GA in CBR revision phase was established in generating timetable solutions [2].The integration of CBR with GA approach has been proved to be better in comparison to other variations of CBR. Nevertheless the application of Genetic algorithms to solve the Cold Start problem occurring in CBR and to generate synchronized and optimized novel plans is still an unexplored research area. In this respect the paper presents modified Genetic algorithm assisted Case Based Reasoning approach for solving Cold-Start problem faced by CBR based systems and generate novel plans. The effectiveness of the approach is established with the help of case study on e-Travel Planning. Proposed approach provides optimized solution plans with minimum duration, synchronized for train and Flight modes of transport within the budget specified by the user. The organization of the paper is as follows. Section 2 presents System architecture and Algorithm. Experiments and results are given in section 3 followed by conclusion and future work in section 4.

2 System Architecture and Algorithm


The architecture of the prototype system developed to demonstrate the capabilities of the evolutionary approach to intelligent planning is shown in the Fig 1. System employs agents to leverage the benefits of Agent architecture [6]. These agents perform the various activities to be carried out by the system. Team of agents include User Agent (UA), Planner Agent (PA), Case Based Reasoning Agent (CBRA), Genetic Algorithm Agent (GAA), Repository Agent (RA) and Verification Agent (VA). The overall algorithm in phases as follows:-

An Evolutionary Approach to Intelligent Planning

307

Fig. 1. System architecture

Phase 1: The User Agent (UA) accepts the query submitted by the user and forwards it to the Planner Agent (PA) for further processing. PA builds an abstract representation of the query received from UA, and processes it to compute solution plans. PA is assisted by two subordinate agents for query processingCase Based Reasoning Agent (CBRA) and Genetic Algorithm Agent (GAA). PA primarily invokes CBRA to find the similar cases based on preferences provided by the user. Phase 2: CBRA predominantly refers the Case Based Reasoning system (CBRs) plan library to check if it has any prior plans stored, with respect to the user specified query. CBRA then tries to fit various modes of transport within the given budget for its optimization. The plans are finally adapted to the new adjusted modes and budget. Phase 3: In this phase the reused plans from the previous phase are revised by changing the cities and budget for further optimization. CBRA passes these plans to PA. PA checks if the number of plans received after revising the cases are sufficient (that is if they are above threshold) to satisfy the user. In such a situation if there are no plans retrieved by CBRA e.g. user submits a new query for which CBR has no past experience, CBR based systems fail to provide any plans to the user. However in our system PA invokes GAA to generate novel plans. Simultaneously PA also invokes VA to verify the current availability status of the plans (if any) generated by CBRA, before communicating them to the UA. It helps in providing only the updated and valid plans to the user e.g. in travel planning domain existing plans may fail if the train or flight no more runs. Phase 4: GAA firstly invokes Repository Agent (RA) to gather route information from the Repository. GAA uses proposed modified genetic algorithm to optimize the route information received from RA and evolve plans within the user specified budget. Steps of the modified Genetic algorithm are as follows:Step 4.1. Size of population determines the number of plans generated by the system. An initial population of 200 chromosomes is generated by GAA. Any change in

308

S. Mehta et al.

population size leads to corresponding change in quantitative number of plans generated by the algorithm. The size and structure of chromosome depends on the user preference of the route. Route with one hop and two hops has Chromosome of size 3 and 4 respectively.
Chromosome Structure for size 3:

Source

Budget

Hop

Budget

Destination

Chromosome Structure for size 4:

Source

Budget

Hop

Budget

Hop

Budget

Destination

Step 4.2: Thereafter a maximum allowable cost is distributed between all two consecutive genes within the chromosome. The sum of all cost within a chromosome is equal to the budget given by the user. Step 4.3: To compute the Fitness of a particular chromosome, the fitness function is as follows:F (Chr) = (n)

(1)

Where Chr is an individual chromosome, n is the total number of ways one can reach the destination from the source.
Step 4.4: The chromosome with highest fitness value is considered most fit. Step 4.5: GAA uses Elitism [7] mechanism for selecting the chromosomes for next generation and performs multipoint crossover with crossover probability of .7 and a mutation probability of 0.03. Step 4.6: GAA employs an additional Evolve operator along with the conventional Crossover and Mutation operators, for optimizing the plans within specified number of generations. The functionality of evolve operator involves duplication of best fit genes with only a slight modification to the parameters. If the new genes are more fit, then the old genes are replaced with the new genes. Thus the genes evolve into better solutions, thereby improving the fitness of the chromosome. Afterwards GAA communicates optimized plans to the Planner Agent. Phase 5: Thereafter GAA stores new plans and their sub-plans in CBRs plan library for solving future queries. Phase 6: PA sends integrated plans (from VA and GAA) to UA to be presented to the user. UA presents the top N plans with minimum time within the budget preferred by the user.

An Evolutionary Approach to Intelligent Planning

309

3 Experiments and Results


Rigorous experiments were performed to assess the competence of proposed approach as compared to traditional CBR system. The prototype system was developed using JADE (Java Agent Development Environment), J2EE (Java 2 Enterprise Edition) and MySql database and executed on Core2 Duo 1.67 GHz processor and 2-GB RAM computer. Specialized programs were developed to retrieve the real datasets of Indian flight [12] and train [13] modes of transport for performing the experiments. The number of stops considered in each case (i.e. flight and train) was different, 162 cities were considered with respect to trains and 20 cities were taken for the flights. All the possible trains and flights between two stops were stored in the database and included in the dataset. On the whole dataset constituted 90000 tuples of various trains and 900 tuples for diverse flights between two stations. The various attributes of trains and flights database include source, destination, fares of different classes, arrival time and departure time.
Experiment 1: This experiment substantiates the ability of system to serve diverse requirements of the users. For example for a query like Source: Bangalore, Destination: Calcutta, Budget: Rs. 9000 and Mode: both (train and flight), algorithm generates varied kinds of plans like 1-hop plans, plans with 2-hops and direct plans as shown in Fig 2

Fig. 2. Optimized and synchronized plans

All these plans are synchronized across train and flight modes of transport with different permutations and combinations as seen in Fig 2 like plans with 2-hops may be FlightFlight-Train or Flight-Train-Train etc. The plans optimized within the budget provided by the user are depicted in the decreasing order of time for complete journey.
Experiment 2: To establish the capability of our approach in solving cold-start problem and generate novel plans, system was tested on 50 random queries with

310

S. Mehta et al.

different source and destinations for both (train and flight) modes of transport and travel budget of around Rs.4000. Fig 3 shows the number of plans with 1-hop and Fig4 exhibits the number of plans with 2-hops between source and destination, generated by CBR alone and CBR-GA hybrid. These figures precisely illustrate that CBR-GA is able to generate plans even for the preliminary queries whereas CBR based system starts generating plans only after having some experience of the related queries. Thus the proposed system is competent in providing results irrespective of any past experience. Fig 5 portrays the total number of plans with zero hops, one hop and two hops between the source and destination retrieved by traditional CBR system and hybrid CBR-GA. These results establish the performance of the system in generating total number of indirect plans (1-hop and 2-hop) as compared to the number of available direct plans. However most of the existing commercial web based travel planning systems present only direct plans to the user which are very less as compared to the number of indirect plans. Together the direct and indirect plans may lead to an enhanced user experience.
Travel Plans with 1-hop
80 70 60 50 40 30 20 10 0

No of Plans

CBR CBR-GA

Number of Plans

qu er y qu 1 er y qu 4 er y7 qu er y qu 10 er y qu 13 er y1 qu 6 er y qu 19 er y qu 22 er y2 qu 5 er y qu 28 er y qu 31 er y qu 34 er y3 qu 7 er y qu 40 er y4 qu 3 er y qu 46 er y4 9
Queries

Fig. 3. Plans with 1-hop

Travel Plans with 2-hops


1600 1400 1200 1000 800 600 400 200 0

CBR CBR-GA

qu er y qu 1 er y4 qu e qu ry 7 er y qu 10 er y1 qu 3 er y qu 16 er y qu 19 er y2 qu 2 er y qu 25 er y2 qu 8 er y qu 31 er y qu 34 er y3 qu 7 er y qu 40 er y4 qu 3 er y qu 46 er y4 9
Queries

Fig. 4. Plans with 2-hop

An Evolutionary Approach to Intelligent Planning

311

Number of Direct Travel Plans and Plans with 1-hop and 2-hop generated by CBR and CBR-GA
4000 3500 3000 2500 2000 1500 1000 500 0
ry qu 1 er y qu 4 er qu y 7 er y qu 10 er y qu 13 er y qu 16 er y qu 19 er y qu 22 er y qu 25 er y qu 28 er y qu 31 er y qu 34 er y qu 37 er y qu 40 er y qu 43 er y qu 46 er y4 9 qu e

Number of Plans

CBR-GA CBR direct

Queries

Fig. 5. Total indirect plans generated by CBR alone and by CBR-GA hybrid as compared to number of direct plans

Experiment 3: It was conducted to show the performance of modified GA as compared to conventional GA as shown in Fig 6. Since the fitness of the chromosome is computed using equation (1), behavior of GA was observed over the 50 different queries. The figure illustrates that over the generations additional evolve operator improves the performance of GA significantly as compared to traditional GA.

Performance comparison
120000 100000 Fitness 80000 60000 40000 20000 0 0 10 20 30 40 50 60 70 Generations Modified GA Conventional GA

Fig. 6. Performance of Modified GA

Experiment 4: This experiment corroborates the ability of system in generating wide variety of optimized plans within a particular cost range. Fig 7 depicts the number of plans generated within a cost range after optimizing the user specified budget (Rs 9000 in this case) across different modes of transport. These plans may help in serving the wide variety of users in a better way.

312

S. Mehta et al.

Number of Plans Generated after Cost Optimization


160 140 120 100 80 60 40 20 0 2000-2999 3000-3999 4000-4999 5000-5999 6000-6999 7000 & Above

Number of Plans

Plans

Cost Range

Fig. 7. Number of plans retrieved within a particular cost range

On the whole, contribution of proposed approach to the existing research is noteworthy. Results of the meticulous experiments evidently demonstrate that proposed approach to CBR-GA hybrid is considerably better than CBR alone. The competency of the system in solving Cold-Start problems and in generating novel plans has been well established through experiments. This approach also curtails the need of maintaining initial case base repository to handle preliminary queries. The ability of modified Genetic Algorithm in cost optimization reveals that large number of plans computed within various ranges may provide better experience to the user. The proposed approach to hybridization of GA with CBR for optimizing and synchronizing plans across multiple modes may be extended for Time Optimization also.

4 Conclusion
This paper presented an intelligent planning approach based on adapted Genetic algorithm (GA) assisted Case Based Reasoning (CBR) to tackle the problems posed in CBR based systems. The efficacy of the approach is demonstrated via case study on e-travel planning for generating the optimized and synchronized indirect travel plans across various modes of transport. Experimental results involving real datasets of Indian train and flight modes of transportation revealed that the algorithm designed using proposed hybrid approach is highly efficient in generating novel plans and in solving cold-start problem stirring in systems developed using CBR alone. Future work includes extending the work to other application domains.

References
[1] Camacho, D., Aler, R., Borrajo, D., Molina, M.J.: Multi-agent plan based information gathering. Journal of Applied Intelligence 25, 5971 (2006) [2] Grech, A., Main, J.: A Case-Based Reasoning Approach to Formulating University Timetables Using Genetic Algorithms. In: Khosla, R., Howlett, R.J., Jain, L.C. (eds.) KES 2005. LNCS (LNAI), vol. 3681, pp. 7683. Springer, Heidelberg (2005)

An Evolutionary Approach to Intelligent Planning

313

[3] Yang, H.-L., Cheng-Shu, W.: Two stages of case-based reasoning - Integrating genetic algorithm with data mining mechanism. Expert Systems with Applications: An International Journal archive 35(1-2), 262272 (2008) [4] Kolodner, J.: Case-Based Reasoning. Morgan and Kaufmann publishers, San Francisco [5] Mitchell, M.: An Introduction to Genetic Algorithm. MIT Press, Cambridge (1999) [6] Wooldridge, M.: An Introduction to Multi Agents. John Wiley & Sons, Chichester (2003) [7] Ahn, H., Kim, K., Han, I.: Hybrid Genetic Algorithms and Case-Based Reasoning Systems. In: International Symposium of Computational and Information Science, pp. 922927. Springer, Heidelberg (2004) [8] Jarmulak, J., Craw, S., Rowe, R.: Genetic Algorithms to Optimise CBR Retrieval. In: Blanzieri, E., Portinale, L. (eds.) EWCBR 2000. LNCS (LNAI), vol. 1898, pp. 159194. Springer, Heidelberg (2000) [9] Mehta, M., Ontan, S., Ram, A.: Using Meta-reasoning to Improve the Performance of Case-Based Planning. In: McGinty, L., Wilson, D.C. (eds.) 8th International Conference on Case Based Reasoning. LNCS, vol. 5650, pp. 210224. Springer, Heidelberg (2009) [10] Niknafs, A.A., Shiri, M.E., Javidi, M.M.: A Case-Based Reasoning Approach in ETourism: Tour Itinerary Planning. In: Proceedings of the 14th International Workshop on Database and Expert Systems Applications. IEEE, Los Alamitos (2003) [11] Shin, K., Han, I.: A Hybrid Approach Using CBR and GA for Corporate Bond Rating. In: International Conference on Case Based Reasoning. ACM, New York (2008) [12] http://www.cleartrip.com [13] http://www.indianrail.gov.in

You might also like