Although there are many sources of information available to a website from which your browser can be identified, the ones listed on the panopticlick experiment provide a good starting point for reducing the size of your browser's fingerprint. Focus on the characterstics that contain the most bits of identifying information.
Another large source of entropy could be the user agent. Not using an exotic browser version or operating system would help, but may be inconvenient. You can also set the user agent manually, from the browser settings or through plugins, to gibberish, although you may be uniquely identified by that, or to a more common string. Shortly after the paniopticlick website was launched, the most frequent user agent in the project's dataset was 'Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv:126.96.36.199) Gecko/20091221 Firefox/3.5.7', but that information is both skewed and outdated. Ideally, you'd have to find more recent data elsewhere.
The HTTP_ACCEPT headers are far less customisable, but you may be able to reduce the entropy a little bit by keeping the number of preferred languages to a minimum or by switching away from an uncommon language. These settings are usually accessible from within the browser. Many websites are only available in one language and will ignore these settings anyway.
Note that these measures don't prevent tracking altogether; they merely reduce the size of your browser's fingerprint. IP addresses, cookies etc. can still be used as identifying information.