In recent years, the use of autonomous agents for pricing products and services has increased rapidly. With the development of machine learning and artificial intelligence agents, agents can actively learn pricing strategies. This allows firms to learn pricing strategies without communicating with other firms with which they should be competing. As collusions can occur even unintentionally and because the current antitrust law framework regulates only intentional communication leading to collusion, existing legal regulations are insufficient in cases where agents automatically learn collusive behavior without exchanging information. To deal with such problems and to promote the future development of pricing by agents, the kinds of algorithms that can make agents collude more strongly must be clarified. Existing studies have shown that reconstructing state representation during training leads to stronger collusion. Therefore, we evaluate a combination of multiple state representations using other statistics and clarify the one leading to more vigorous collusion. We clarify that the state with the minimum and the second minimum leads the significantly stronger collusion in a three-company competition than in the conventional case by about 4.25%. We also analyze postlearning agent strategies and the effectiveness of the state representation of the strategy.